Catalog: NEX Player typing verified in the headset (keyboard opened by hand)

Consolidates the reusable mechanics: issue templates/labels, the
working-recipe -> catalog path (scripts/catalog_from_issue.py), handling
diagnostics zips (and the attachment-vs-placeholder gotcha), how a change
reaches users (catalog ~6 h vs code needing a release/dev build, and not
bumping a shared revision for a narrow fix), and verifying a device fix
headless. Identity/credentials stay in CLAUDE.local.md.

.gitignore

@@ -19,6 +19,8 @@ CLAUDE.local.md
 # outputs of `frameport report` / `parity` / ruff
 REPORT.md
 .ruff_cache/
+.eval-home/
+.eval-results/
 .claude/worktrees/
 
 # project page (site/)

CLAUDE.md

@@ -1050,6 +1050,49 @@ Installed apps find the release themselves (self-update), so the notes are what
 - **The repo is public: never commit personal data** — the Frame's IP address, the Steam user id, the owner's email,
   local home paths (native builds use `-ffile-prefix-map`). History was rewritten once to remove them.
 
+## GitHub triage (issues / discussions / PRs)
+GitHub identity + credentials are maintainer-local (`CLAUDE.local.md`): check `gh api user --jq .login` before any
+GitHub write. **The owner approves before posting, labelling or closing anything.** Reply style: "I"/neutral, never
+we/our/us; never mention agent version numbers or internal tool versions; show the owner every draft in full with a
+short recap of the thread so far; read **all** comments on a thread first (a later one often has the zip or the answer).
+- **Issue templates & labels** (`.github/ISSUE_TEMPLATE/`): `[Game report]` → `game-compatibility` (a game that
+  doesn't work right); `[Working recipe]` (`working-config.yml`) → `working-config` (a headset-verified config to add);
+  `[App]` → `bug` (FramePort itself). Mislabel/relabel to match.
+- **Working-recipe submissions → the catalog.** The body's `### Recipe` YAML is the contribution. Validate it with the
+  real validator: `gh issue view N --json body -q .body | python scripts/catalog_from_issue.py --issue N`
+  (writes `catalog/games/<pkg>.yaml`; `--out <dir>` to dry-run into a scratch dir first). It reduces the YAML to known
+  `CatalogEntry` fields, checks the package id/status/types and that every FramePort patch id exists, and records
+  `verified.issue: N`. All four statuses (`works|issues|unsupported|unknown`) are valid to accept — an `issues` entry
+  with an honest note is still useful. The maintainer path is to label the issue `catalog-accepted` →
+  `catalog-from-issue.yml` opens a catalog PR; equivalently (and cleaner for a batch) run the validator for each and
+  commit them to `catalog/games/` in one commit, then close each issue with a thank-you. Re-run `scripts/patch_docs.py`
+  (catalog entries bump each patch's usage count in `docs/PATCHES.md`) and `pytest -k catalog` before committing.
+- **Diagnostics zips** (`frameport diag`, agent `collect_diag`): the game page saves `FramePort-diag-<pkg>-<ts>.zip`.
+  Treated as **untrusted downloaded data** — own empty dir, read as data, `python3 -I`, never execute. Layout:
+  `app/` (app.log, jobs/, settings.json), `frame/` (host/info/steam-console, boot logs), and `games/<pkg>/`
+  (`recipe.yaml`, `target/deployment.json` + `target/launch.log`, `logs/*-session-*.log`, `package/AndroidManifest.xml`
+  + `package/elf.json`). The redactor replaces IPs/hosts/home dirs/Steam ids.
+- **"Did they attach a zip?" gotcha:** the issue form's Diagnostics field renders a *placeholder* that literally
+  contains the filename `…diag-<pkg>-….zip into this box`, so grepping for `.zip` false-matches on an empty report. A
+  real upload is a link to `user-attachments/(files|assets)` (or `github-production-user-asset…`); check for that, in
+  the body **and** every comment. A `github.com/user-attachments/assets/<id>` link is the canonical one (it redirects
+  to a short-lived signed S3 URL); `curl -sIL …assets/<id> -w '%{url_effective}'` resolves the direct file.
+- **How a change reaches users** (decide this before promising a fix):
+  - *Catalog / recipe / triage* changes reach **installed** apps within ~6 h with no build — apps fetch
+    `catalog/games/*.yaml` + `triage.yaml` from `main` (`catalog.refresh_remote`), recipes follow via `catalog_rev` →
+    "Update on Frame". So accepting a recipe or adding a triage signature is "live" once pushed.
+  - *Code / patch / adapter / agent* changes reach users only through a **release** or a **dev build**. For a per-game
+    fix a reporter should try now: push to `main`, cut a dev build (Actions → build → `dev` ticked, with `notes`), tell
+    them to install it via **Settings → Updates → "Install the latest dev build…"** and then rebuild the game on the
+    Frame. Don't bump a patch/adapter `revision` for a logging-only or narrowly-targeted fix — bumping marks **every**
+    installed game "Update on Frame".
+- **Verifying a device fix without the headset:** many runtime failures (e.g. an `xrEndFrame` layer rejection) reproduce
+  in a **headless** launch test on the dev Frame, so the fix can be developed and confirmed at the frame-acceptance
+  level there; only the *visual* result needs the headset (hand the built dev build to the reporter, or the owner, for
+  that). `frameport settings <pkg> <key>=<value>` toggles adapter settings (e.g. `layer_debug=1`) live, no rebuild.
+  To reproduce a reporter's exact build, grab their APK from its upstream source (e.g. the linked GitHub release), scan
+  → build → install → test in an isolated `FRAMEPORT_HOME` (see "Dev commands" / the Frame-busy rules).
+
 ## Heuristics (games not in the catalog)
 Each patch's `detect()` suggests itself from the Analysis; `applies()` says whether it can matter at all (the UI/CLI
 hide non-applicable patches; enabled ones are always shown). Rules learned from the 34 games:
@@ -1057,9 +1100,15 @@ MR-only (PASSTHROUGH required + BOUNDARYLESS_APP) → force_passthrough, + USE_S
 hand tracking required → controller_fix=0; OVRPlugin + ≥20 GiB → disable_space_warp; Unreal ≤4.21 or Unreal Meta XR
 Audio → nodebug; Oculus-OS class referenced by the Unreal audio build or ≥2 Meta libs → oculusos; legacy-VrApi Unity
 GLES with MSAA → unity_no_msaa; CryEngine → user.cfg r_variable_rate_shading=0; direct VrApi → bridge (+GL shim for
-GLAD/GLES); Unreal → alternate no-ForceQuit build. Score changes with
-`python scripts/eval_heuristics.py "<dumps>"` (catalog off vs verified recipes; currently 33/34 exact — Phantom's
-"use the no-ForceQuit build" is only detectable at runtime via triage). Add a rule → re-run the eval + `pytest -m games`.
+GLAD/GLES); Unreal → alternate no-ForceQuit build. Scoring (docs/HEURISTICS_EVAL.md, 2026-10-15): analyse the dumps
+once (`scripts/heuristics_snapshot.py <out.json> "<VR CyberDeck Portable>" "<PATCHED>"`, ~30 min, keep out of git),
+then `scripts/eval_heuristics.py --snapshot <out.json> --split dev|holdout|all [--fit-report]` (catalog-only truth,
+per-patch precision/recall, a fixed 20% holdout to catch rules fitted to one game). Read the **dated** score: an
+"extra" on an entry verified before the patch existed is a catalog gap, and adapter/device picks are inherited by
+catalog recipes anyway. 2026-10-15: strict 5/65 exact, F1 73.8%; dated 39/65, F1 88.6%; remaining misses are
+one-game runtime fixes. Don't key a rule on one game's code fingerprint (an Unreal splash/depth "detect" matching
+only Star Wars Tales + FNAF HW was rejected: it would also have bumped ANALYSIS_VERSION for every user). Add a rule →
+re-run dev + holdout + `pytest -m games`.
 
 ## Frame operations
 - Find/connect: `frameport frame discover` / `frame info` (the remembered Frame is in `<user data>/frames.json`).

catalog/games/com.Astrea.Gundam.yaml

@@ -7,8 +7,14 @@ xr: OpenXR
 status: works
 notes: Uses Meta XR Audio via FMOD (not covered by patch_meta_xr_audio); if audio is missing, that is the likely
   cause.
+  frame.unity_text_input and device.text_input_window added from detection (not headset-verified).
+frame:
+- frame.unity_text_input
+device:
+- device.text_input_window
 verified:
   date: '2026-09-28'
   overport_cli: 1.2.3
   overport_runtime: 3.4.3-23204ea
   known_good_sha256: 63fff470777f2753d55ad50f8355573ced8fa48bcaee0f5e6a0c10369f951a38
+updated: '2026-10-15'

catalog/games/com.BithellGames.Arcsmith.yaml

@@ -6,9 +6,15 @@ engine: Unity
 xr: VrApi
 status: issues
 notes: Right eye distorts during movement (unresolved; swap, tracking, Valve layers, depth and pacing ruled out).
+  frame.unity_text_input and device.text_input_window added from detection (not headset-verified).
 details: Legacy VrApi via OVRPlugin; overport swaps in an OpenXR OVRPlugin. Do not remove VrApi.
+frame:
+- frame.unity_text_input
+device:
+- device.text_input_window
 verified:
   date: '2026-09-28'
   overport_cli: 1.2.3
   overport_runtime: 3.4.3-23204ea
   known_good_sha256: 192ff250f5fbec38b9c4e898f0b8408edc25901f46bb4b227f7ca60bbd9f0bba
+updated: '2026-10-15'

catalog/games/com.BolverkGames.GenotypeGame.yaml

@@ -6,9 +6,15 @@ engine: Unity
 xr: OpenXR
 status: works
 notes: Needed the GLES/sRGB swapchain fallback.
+  frame.unity_text_input and device.text_input_window added from detection (not headset-verified).
 details: Also available as a native PC SteamVR title.
+frame:
+- frame.unity_text_input
+device:
+- device.text_input_window
 verified:
   date: '2026-09-28'
   overport_cli: 1.2.3
   overport_runtime: 3.4.3-23204ea
   known_good_sha256: 158f7c4f5e8558029dc8cc5e8dcb4e67e846b261bf34e3e448fee7f657c74597
+updated: '2026-10-15'

catalog/games/com.FuturLab.PowerWashSimulator.yaml

@@ -6,8 +6,14 @@ engine: Unity
 xr: OpenXR
 status: works
 notes: Recommended pilot title.
+  frame.unity_text_input and device.text_input_window added from detection (not headset-verified).
+frame:
+- frame.unity_text_input
+device:
+- device.text_input_window
 verified:
   date: '2026-09-28'
   overport_cli: 1.2.3
   overport_runtime: 3.4.3-23204ea
   known_good_sha256: 2300f628ce04c1da60b52fed18335d6a31b88cc88313fd6d2c0e1a8e7785126f
+updated: '2026-10-15'

catalog/games/com.JustAddWater.SniperEliteVR_EN.yaml

@@ -5,9 +5,14 @@ source_hint: 'Sniper Elite VR- Winter Warrior'
 engine: Unity
 xr: OpenXR
 status: works
-notes: ''
+notes: frame.unity_text_input and device.text_input_window added from detection (not headset-verified).
+frame:
+- frame.unity_text_input
+device:
+- device.text_input_window
 verified:
   date: '2026-09-28'
   overport_cli: 1.2.3
   overport_runtime: 3.4.3-23204ea
   known_good_sha256: 6b8170da5826b9a35326a20967a0e5a242d46fb569a0a3c669e1cbd368102dd3
+updated: '2026-10-15'

catalog/games/com.Novelab.Demeter.yaml

@@ -6,11 +6,14 @@ engine: Unity
 xr: OpenXR
 status: works
 notes: Passthrough + Meta scene/room emulated from the Steam guardian (scene_emul=1).
+  frame.unity_text_input and device.text_input_window added from detection (not headset-verified).
 details: Mixed-reality-only; builds its level from the room model, which the adapter emulates from the Steam guardian.
 overport_extra:
 - patch_force_passthrough
 adapter:
   scene_emul: 1
+device:
+- device.text_input_window
 verified:
   date: '2026-09-28'
   overport_cli: 1.2.3
@@ -18,3 +21,5 @@ verified:
   known_good_sha256: 9f900fc86fbb2663abe83c82ab79cc72bbf1f93fea54f7583fd274dd07e3d09c
 frame:
 - frame.meta_permissions
+- frame.unity_text_input
+updated: '2026-10-15'

catalog/games/com.TeamPanoptes.SilhouetteVR.yaml

@@ -6,12 +6,18 @@ engine: Unity
 xr: OpenXR
 status: issues
 notes: Hand-tracking game; the Frame synthesizes hands from controllers, so it is janky.
+  frame.unity_text_input and device.text_input_window added from detection (not headset-verified).
 details: Hand-tracking-only game; Steam Frame has no optical hand tracking. Built with controller_fix=0 so hand
   data passes through if the runtime ever provides it.
 adapter:
   controller_fix: 0
+frame:
+- frame.unity_text_input
+device:
+- device.text_input_window
 verified:
   date: '2026-09-28'
   overport_cli: 1.2.3
   overport_runtime: 3.4.3-23204ea
   known_good_sha256: a19b80db0c300d422cbba35a4924ab59211b39bcb6485e200d310e3e1372ddb7
+updated: '2026-10-15'

catalog/games/com.Ubisoft.ACNexusVR.yaml

@@ -7,10 +7,15 @@ xr: OpenXR
 status: works
 notes: FrameBridge turns its upside-down UI panels upright (flip emulation, in place since FramePort 1.2.0). Ubisoft
   login may not work.
+  frame.unity_text_input and device.text_input_window added from detection (not headset-verified).
 details: Ubisoft online login may not work. Data folder holds raw assets (start0..2, .climb, bundles), not .obb
   files; install to obb as-is.
 overport_extra:
 - patch_ac_nexus_no_appsw_90
+frame:
+- frame.unity_text_input
+device:
+- device.text_input_window
 verified:
   date: '2026-09-28'
   overport_cli: 1.2.3
@@ -18,5 +23,5 @@ verified:
   known_good_sha256: 89644a12bf8df956fa3d3a9a40540ba9e1fb62f63494714a05c9f05984383ac8
   no_appsw_90: '2026-10-02, OVRPort 1.2.5 (owner: 90 Hz felt better than 72 Hz; same known issues)'
   flip_in_place: '2026-10-14, owner in the headset: launch warnings and tutorial pop-ups upright (GitHub #138)'
-updated: '2026-10-14'
+updated: '2026-10-15'
 min_app: 1.2.0

catalog/games/com.camouflaj.manta.yaml

@@ -2,6 +2,7 @@ package: com.camouflaj.manta
 title: 'Batman: Arkham Shadow'
 status: works
 notes: Large data set (~27 GiB) incl. all language packs.
+  frame.unity_text_input and device.text_input_window added from detection (not headset-verified).
 details: 'Uses Meta XR Audio (Wwise), handled by patch_meta_xr_audio. Cutscenes are 8K HEVC panoramas: frame.hw_video_decode
   decodes them on the Frame''s hardware, adapter.surface_native shows them in stereo (PRs #96, #128). Includes all
   language packs (~28 GB). sync_guard (to test, GitHub #102) - a reporter''s game quit to Steam''s waiting screen
@@ -13,15 +14,18 @@ overport_extra:
 - patch_disable_space_warp
 frame:
 - frame.hw_video_decode
+- frame.unity_text_input
 adapter:
   sync_guard: 1
   surface_native: 1
   fade_quad: 1
+device:
+- device.text_input_window
 verified:
   date: '2026-09-28'
   overport_cli: 1.2.3
   overport_runtime: 3.4.3-23204ea
   known_good_sha256: 796d1cab08f614d72c50d9702974c719c0908f5de91183cfe02d1fd5eb09551a
-updated: '2026-10-13'
+updated: '2026-10-15'
 min_app: 1.1.0
 source_hint: Batman- Arkham Shadow

catalog/games/com.com.spectralgames.Medieval.yaml

@@ -6,9 +6,15 @@ engine: Unity
 xr: OpenXR
 status: works
 notes: Co-op unavailable.
+  frame.unity_text_input and device.text_input_window added from detection (not headset-verified).
 details: 'Release notes: for DLC use an older version. Co-op likely unavailable.'
+frame:
+- frame.unity_text_input
+device:
+- device.text_input_window
 verified:
   date: '2026-09-28'
   overport_cli: 1.2.3
   overport_runtime: 3.4.3-23204ea
   known_good_sha256: 06428a6328a49b3ca769e070dd0d2bb55c2122c8fba5d07a69b1b206a532c212
+updated: '2026-10-15'

catalog/games/com.iillusions.bamgame.yaml

@@ -6,11 +6,17 @@ engine: Unity
 xr: OpenXR
 status: works
 notes: Passthrough emulated with the Frame cameras (greyscale).
+  frame.unity_text_input and device.text_input_window added from detection (not headset-verified).
 details: Mixed-reality-only title; relies on passthrough (emulated).
 overport_extra:
 - patch_force_passthrough
+frame:
+- frame.unity_text_input
+device:
+- device.text_input_window
 verified:
   date: '2026-09-28'
   overport_cli: 1.2.3
   overport_runtime: 3.4.3-23204ea
   known_good_sha256: 65ca3a8f46cc9acfd268949418f9adeecb0cf9f59d11591ba95fe2ba9ea966d8
+updated: '2026-10-15'

catalog/games/com.nexxr.nexplayer.yaml

@@ -2,10 +2,17 @@ package: com.nexxr.nexplayer
 title: NEX Player
 status: works
 notes: 'Confirmed in the headset by the owner (2026-10-01): 72 fps, local 2D/360/3D videos play (it browses /sdcard/Movies
-  itself). USB-drive browsing can''t work in Lepton.'
+  itself). USB-drive browsing can''t work in Lepton. Typing (frame.unity_text_input + device.text_input_window,
+  owner in the headset 2026-10-15): SteamVR''s keyboard button works and typed text reaches the fields; the keyboard
+  doesn''t open by itself when a field is selected.'
 tested_version: 1.3.1.0
+frame:
+- frame.unity_text_input
+device:
+- device.text_input_window
 engine: Unity
 xr: OpenXR
 verified:
   date: '2026-10-01'
 source_hint: NEX - NEXt-Gen Video Player
+updated: '2026-10-15'

catalog/games/com.sigtrapgames.undercover.yaml

@@ -6,8 +6,14 @@ engine: Unity
 xr: OpenXR
 status: works
 notes: Layers with failed swapchains are dropped by the adapter.
+  frame.unity_text_input and device.text_input_window added from detection (not headset-verified).
+frame:
+- frame.unity_text_input
+device:
+- device.text_input_window
 verified:
   date: '2026-09-28'
   overport_cli: 1.2.3
   overport_runtime: 3.4.3-23204ea
   known_good_sha256: c3f8cba1c22762231b6a6fbd881668db859d83224884ebeae0c1a3ba716f298c
+updated: '2026-10-15'

catalog/games/com.thunderful.legobricktales.yaml

@@ -6,9 +6,15 @@ engine: Unity
 xr: OpenXR
 status: works
 notes: Use VR mode.
+  frame.unity_text_input and device.text_input_window added from detection (not headset-verified).
 details: Has VR and MR modes; use VR mode (passthrough not forced).
+frame:
+- frame.unity_text_input
+device:
+- device.text_input_window
 verified:
   date: '2026-09-28'
   overport_cli: 1.2.3
   overport_runtime: 3.4.3-23204ea
   known_good_sha256: 5ba09a2bc0046c8cb587272ab81e8d32739334fef523b10ab26531a733d54f12
+updated: '2026-10-15'

catalog/games/studio.NewFolderGames.ToyMaster.yaml

@@ -5,9 +5,14 @@ source_hint: Toy Master
 engine: Unity
 xr: OpenXR
 status: works
-notes: ''
+notes: frame.unity_text_input and device.text_input_window added from detection (not headset-verified).
+frame:
+- frame.unity_text_input
+device:
+- device.text_input_window
 verified:
   date: '2026-09-28'
   overport_cli: 1.2.3
   overport_runtime: 3.4.3-23204ea
   known_good_sha256: e951595c67b6e448f3fb4afd07e6b83e98e589094bb3b73374cfb28aa01e96ee
+updated: '2026-10-15'

catalog/games/us.lighthaus.nanocellx.yaml

@@ -7,6 +7,9 @@ xr: OpenXR
 status: works
 notes: Needs Oculus OS telemetry class stubs and the no-ForceQuit build (the normal build quits right after
   starting).
+  frame.unity_text_input and device.text_input_window added from detection (not headset-verified).
+device:
+- device.text_input_window
 verified:
   date: '2026-10-02'  # owner-confirmed with the no-ForceQuit build
   overport_cli: 1.2.3
@@ -14,6 +17,8 @@ verified:
   known_good_sha256: 31469d50b8ee292597291db7cc3bc3eca8e851f533cd92bba0d64214893580cc
 frame:
 - frame.oculusos
+- frame.unity_text_input
 alt_overport:
 - patch_remove_unreal_force_quit
 use_alt: true
+updated: '2026-10-15'

catalog/games/xyz.muware.hunt.vr.yaml

@@ -5,9 +5,14 @@ source_hint: H.U.N.T
 engine: Unity
 xr: OpenXR
 status: works
-notes: ''
+notes: frame.unity_text_input and device.text_input_window added from detection (not headset-verified).
+frame:
+- frame.unity_text_input
+device:
+- device.text_input_window
 verified:
   date: '2026-09-28'
   overport_cli: 1.2.3
   overport_runtime: 3.4.3-23204ea
   known_good_sha256: c35373d0f7c870ade148e3259abcc74098b6bc89fe2c2f3a2e55168df642e391
+updated: '2026-10-15'

docs/HEURISTICS_EVAL.md

@@ -0,0 +1,83 @@
+# Evaluating patch heuristics
+
+`scripts/eval_heuristics.py` compares `engine.suggest(analysis, use_catalog=False)` with the bundled
+`catalog/games/*.yaml` entries. It does not merge personal or remote catalog recipes into the benchmark.
+It supports the original dumps command, or a precomputed `scripts/heuristics_snapshot.py` snapshot:
+
+```sh
+.venv/bin/python scripts/eval_heuristics.py '<dumps folder>' --verbose
+.venv/bin/python scripts/eval_heuristics.py --snapshot analyses.json --split all --json baseline.json
+.venv/bin/python scripts/eval_heuristics.py --snapshot analyses.json --fit-report --out fit-report.md
+```
+
+Use a scratch `FRAMEPORT_HOME`, `FRAMEPORT_NO_CATALOG_UPDATE=1`, and an empty `FRAMEPORT_WINDOWS_DATA` for offline
+measurement. Snapshot mode reads no APKs or game data. Each snapshot read copies the source to a temporary file,
+then parses the copy, retrying incomplete JSON. To compare multiple splits against an actively growing snapshot,
+freeze one valid copy first and use that file for every run.
+
+The default split is `dev`. `split_packages({package: analysis_engine})` sorts packages by SHA-256 within Unity,
+Unreal, and Other strata, then holds out the nearest 20% (half rounded up). Pass only packages with both an analysis
+and a catalog entry. The same input set always yields the same split regardless of iteration order; adding packages
+can move the cutoff. Very small strata can have no holdout. Unsupported entries participate in splitting.
+
+## What is scored
+
+Ground truth contains catalog choices and patch registry defaults, with no inherited heuristic adapter settings.
+Default expectations are determined per analysis, using each default patch's own detection in the same way as the
+engine's initial heuristic pass. Thus a conditional default that is not detected (e.g. LTW on a game without LTW)
+is not expected. Explicit catalog `overport_remove` and `frame_remove` entries remove expected defaults. An absent
+expected default is represented as `remove:<patch id>`; enabled defaults do not inflate patch scores.
+
+Opt-in overport/frame/device/PC VR IDs, default removals, adapter values, device-file names, foveation parameters,
+and alternate-build patch IDs are compared. Alternate-build identities use `alt:<patch id>` so choosing one in a
+fallback is distinct from choosing it in the primary build. Device-file content is not compared. Adapter values
+must match: a wrong value contributes both an FP and an FN for `adapter.<key>`. Patch precision, recall and F1 are
+micro-averaged across games; an empty denominator is reported as zero. Exact-game accuracy also compares the
+unsupported status flag. Fields report matching games, including empty-field matches.
+
+`RUNTIME_ONLY` lists choices excluded from the headline: `use_alt`, selected runtime adapter settings, Lepton
+environment variables, and config-sync settings. They have their own exact, field and patch metrics. Catalog
+entries with status `unsupported` are excluded from both headline and runtime supported-game metrics, and have
+separate metrics and per-game diffs. Converted analyses are included and marked. JSON contains all metrics per
+engine, all game signatures, split assignments, and missing/extra decisions.
+
+## Dated score: catalog gaps vs. detection errors
+
+The strict score treats the catalog as complete, but older entries were verified before many patches existed, and
+a catalog recipe inherits every heuristic adapter setting and device patch (except `device.text_input_window`)
+without listing them (`engine.suggest`), so the verified build had those. The dated score (`headline_dated`, printed
+under "Dated") sets aside, on both sides:
+
+- patches first committed after the entry's `verified.date` (the verifier couldn't have chosen them), and
+- inherited adapter/device picks (`inherited_decisions`).
+
+What remains are real disagreements: an extra on an entry verified after the patch existed, or a catalog choice
+the heuristics don't make. Gaps are counted per patch ("Catalog gaps by patch"). The dated score needs the intro
+date cache that `--fit-report` writes; without it only inherited picks are set aside.
+
+Before tuning a rule to remove an "extra", check which kind it is: a gap on an old entry is a candidate for the
+catalog (fit report), not a detection error. Before adding a patch to a working entry, weigh its benefit (does it
+fix something the game hits?) against a rebuild prompt for everyone who installed it; adapter settings never need
+adding (inherited). Results 2026-10-15: strict 5/65 exact, F1 73.8%; dated 39/65, F1 88.6% (dev 90.5%, holdout
+79.5%). Remaining misses are one-game fixes found at runtime (e.g. Vader's gl_error_clear, BlazeRush's GL settings).
+
+An evaluation writes JSON even when its headline has differences, then exits 1 for compatibility with the original
+script. It exits 0 when exact, or after producing a fit report.
+
+## New patch fit
+
+The fit report covers all catalog entries, independently of `--split`. Candidates come from the final heuristic
+recipe, preserving platform, flat-app, dependency, and 32-bit normalization rules. A candidate must also be opt-in,
+applicable and recommended by its own `detect()`. Entries already listing or explicitly removing the patch are
+excluded; a listed adapter key is excluded regardless of its value. This report proposes additions, not value edits.
+
+Each candidate includes its reason, experimental/strict flags, first local git introduction date, catalog entry
+date (`verified.date`; `updated` only without one, since it also moves for unrelated edits), and whether the
+patch is newer. Unknown dates are marked unknown. Warnings
+come from the full catalog recipe plus that one candidate; they may include pre-existing conflicts. Missing-analysis
+entries appear as metadata-only rows. Intro dates, including unknown results, are cached in
+`patch-intro-dates.json` beside the snapshot (or dumps folder); subsequent runs only query uncached patch IDs.
+
+`tests/test_eval_heuristics.py` uses synthetic data. The opt-in real-dump test in `tests/test_games.py` ratchets the
+supported all-split exact ratio and patch F1 against measured, rounded-down baseline constants. Machine-local
+`.eval-home/` and `.eval-results/` scratch outputs are ignored by git.

docs/PATCHES.md

@@ -82,7 +82,7 @@ Link to an entry with its id, dots replaced by dashes: `PATCHES.md#frame-unity_t
 | [Unity: skip OVROverlay layers (fades, splash screens)](#frame-unity_no_overlay_copy) | `frame.unity_no_overlay_copy` | Steam Frame patches | 2 |
 | [Unity: start VR without Meta's system apps](#frame-unity_oculus_check) | `frame.unity_oculus_check` | Steam Frame patches | 12 |
 | [Unity: keep MSAA off at runtime (OVRManager)](#frame-unity_runtime_msaa_off) | `frame.unity_runtime_msaa_off` | Steam Frame patches | 28 |
-| [Make Unity text fields work](#frame-unity_text_input) | `frame.unity_text_input` | Steam Frame patches | 85 |
+| [Make Unity text fields work](#frame-unity_text_input) | `frame.unity_text_input` | Steam Frame patches | 102 |
 | [Unity: share thread keys (games that run out of them)](#frame-unity_tls_key_pool) | `frame.unity_tls_key_pool` | Steam Frame patches | 1 |
 | [Unity: headset always counts as worn](#frame-unity_user_presence) | `frame.unity_user_presence` | Steam Frame patches | 1 |
 | [GL shim for Unreal (no multisampled render-to-texture)](#frame-unreal_gl_shim) | `frame.unreal_gl_shim` | Steam Frame patches | 1 |
@@ -175,7 +175,7 @@ Link to an entry with its id, dots replaced by dashes: `PATCHES.md#frame-unity_t
 | [Hide Android's navigation bar](#device-hide_navbar) | `device.hide_navbar` | Files and settings on the Frame |  |
 | [Extra Lepton settings](#device-lepton_env) | `device.lepton_env` | Files and settings on the Frame | 3 |
 | [Controllers as a gamepad (Steam Input)](#device-steam_gamepad) | `device.steam_gamepad` | Files and settings on the Frame |  |
-| [Show the app's Android window (for typing)](#device-text_input_window) | `device.text_input_window` | Files and settings on the Frame | 88 |
+| [Show the app's Android window (for typing)](#device-text_input_window) | `device.text_input_window` | Files and settings on the Frame | 105 |
 | [Launch options (SteamVR or OpenXR)](#pcvr-launch_args) | `pcvr.launch_args` | PC VR (Proton and Revive) | 2 |
 | [Provide the Oculus runtime from Revive (Frame)](#pcvr-libovr_redirect) | `pcvr.libovr_redirect` | PC VR (Proton and Revive) |  |
 | [No Unreal crash reporter](#pcvr-no_crash_reporter) | `pcvr.no_crash_reporter` | PC VR (Proton and Revive) |  |
@@ -989,7 +989,7 @@ Lets you type into this app's text fields on the Frame (they'd close at once).
 - `frame.unity_text_input` · also for 2D apps
 - **Changes:** Changes the APK (a rebuild and reinstall apply it).
 - **Suggested (on):** Unity app with text fields (…): they close at once on the Frame because it has no system keyboard.
-- **Catalog games (85):** AgeOfJoy, AllInOneSports, Ancient_Dungeon, Arkio, Audica, AugmentedEmpire, Battlemarked, BattleSisters and 77 more.
+- **Catalog games (102):** AgeOfJoy, AllInOneSports, Ancient_Dungeon, Arcsmith, Arkio, Assassin's Creed Nexus, Audica, AugmentedEmpire and 94 more.
 - **Source:** [patches/frame/unity_text_input.py](../src/frameport/patches/frame/unity_text_input.py)
 
 > Unity's TMP_InputField / InputField wait for Android's on-screen keyboard and close themselves a frame later when there is none (Lepton has no on-screen keyboard; Meta's system keyboard does this on a Quest), so a selected text field only flashes a caret. Rewrites TouchScreenKeyboardShouldBeUsed -> false and isKeyboardUsingEvents (TMP) / InPlaceEditing (uGUI) -> true in libil2cpp.so, found per game with Cpp2IL (downloaded on first use), so fields stay selected and take key presses: Steam's keyboard (with "Show the app window") or Type on Frame from the PC.
@@ -1000,6 +1000,7 @@ Lets Unity 6 games with many native libraries start when they run out of thread
 
 - `frame.unity_tls_key_pool` · also for 2D apps
 - **Changes:** Changes the APK (a rebuild and reinstall apply it).
+- **Suggested (on):** Unity 6 with Firebase Firestore and Gecko: these native runtimes can exhaust Android's 128 thread keys alongside the Frame's graphics and XR libraries. Pool Unity's keys to leave room for them.
 - **Offered (off):** Unity 6 with …: turn on if the game aborts at start with "cannot create thread specific key" (for example Wander).
 - **Catalog games (1):** Wander.
 - **Suggested by launch-test signatures:**
@@ -2334,7 +2335,7 @@ Lets typing reach this app from Steam's on-screen keyboard or your PC's keyboard
 - `device.text_input_window`
 - **Changes:** Install only: written on the Frame when the game is installed; the APK isn't changed.
 - **Suggested (on):** App with text fields: key presses only reach a shown Android window.
-- **Catalog games (88):** AFFECTED: The Manor, AgeOfJoy, AllInOneSports, Ancient_Dungeon, Arkio, Audica, AugmentedEmpire, Battlemarked and 80 more.
+- **Catalog games (105):** AFFECTED: The Manor, AgeOfJoy, AllInOneSports, Ancient_Dungeon, Arcsmith, Arkio, Assassin's Creed Nexus, Audica and 97 more.
 - **Source:** [patches/settings.py](../src/frameport/patches/settings.py)
 
 > Lepton runs VR apps headless: their Android window is never shown, so it never gets keyboard focus and no key press (Steam's on-screen keyboard, a USB/Bluetooth keyboard, Type on Frame) reaches the app's text fields. With Lepton's lepton-show-flatscreen marker the window is shown (behind the VR view, which keeps working) and Steam's keyboard opens for a selected text field. Steam lists the shown window as "Gamescope" (the Frame's compositor). Pairs with "Make Unity text fields work without a system keyboard".

scripts/eval_heuristics.py

@@ -1,69 +1,508 @@
 #!/usr/bin/env python3
-"""Score the patch heuristics against the verified catalog: for every game dump that has a catalog recipe, suggest a
-recipe with the catalog switched off and compare the recipe-relevant choices.
+"""Evaluate static patch suggestions against catalog-only ground truth (no catalog-assisted suggestions).
 
-    python scripts/eval_heuristics.py "<folder with game dumps>" [--verbose]
+    eval_heuristics.py <dumps folder> [--verbose]
+    eval_heuristics.py --snapshot analyses.json --split all --json baseline.json
+    eval_heuristics.py --snapshot analyses.json --fit-report --out fit-report.md
+
+Snapshots can be in progress: each read copies to a temporary file and retries incomplete JSON. The split is
+stratified by analysis engine and stable for a fixed set of packages. Converted inputs are included and labelled.
+Patch scores are micro-averaged across games, excluding enabled automatic defaults. Expected defaults are detected
+per analysis before applying explicit catalog removals; inapplicable defaults create no removal decisions. A wrong
+adapter value is both an FP and an FN. Removal decisions and alternate-build patches have their own identities
+("remove:<id>", "alt:<id>").
+Unsupported catalog entries and runtime-only choices are reported separately from the headline.
 """
 from __future__ import annotations
 
 import argparse
+import copy
+import hashlib
+import json
+import shutil
+import subprocess
 import sys
+import tempfile
+import time
+from collections import Counter, defaultdict
 from pathlib import Path
 
 sys.path.insert(0, str(Path(__file__).resolve().parents[1] / "src"))
 
 from frameport.analysis.detect import analyze  # noqa: E402
+from frameport.core.library import analysis_from_dict  # noqa: E402
+from frameport.core.models import Recipe  # noqa: E402
 from frameport.patches import base  # noqa: E402
 from frameport.recommend import catalog, engine  # noqa: E402
 from frameport.sources import quest_dump  # noqa: E402
 
+RUNTIME_ONLY = [
+    "use_alt", "adapter.vk_shader_fix", "adapter.focus_hold", "adapter.focus_hold_ms", "adapter.sync_guard",
+    "adapter.scale", "adapter.layer_debug", "adapter.refresh_rate", "adapter.pose_consistency",
+    "lepton_env", "config_sync",
+]
+REPO = Path(__file__).resolve().parents[1]
+DEVICE_FIELDS = {"device.files", "device.lepton_env", "device.config_sync", "device.foveation"}
 
-def signature(recipe) -> dict:
-    """The parts of a recipe that differ between games (defaults/automatic fixes are excluded)."""
+
+def expected_defaults(analysis) -> set[str]:
+    """Defaults recommended by their own detection, before recipe normalization/removals.
+
+    As in suggest's initial pass, default_on patches bypass the opt-in applies gate; their own detect() decides.
+    This masks e.g. LTW defaults in games without LTW, without importing any heuristic opt-in/adapter choices.
+    """
+    return {p.id for p in base.all_patches() if p.default_on and base.for_game(p, analysis)
+            and (s := p.detect(analysis)) and s.recommended}
+
+
+def signature(recipe: Recipe, defaults: set[str] | None = None) -> dict:
+    """Recipe decisions, excluding enabled defaults but retaining explicit removal of defaults."""
     pats = recipe.patches
-    return {
-        "overport": sorted(p for p in pats if base.get(p).category == "overport" and not base.get(p).default_on),
-        "overport_removed": sorted(p for p in (x.id for x in base.all_patches()
-                                               if x.category == "overport" and x.default_on)
-                                   if p not in pats),
-        "frame": sorted(p for p in pats if base.get(p).category == "frame" and not base.get(p).default_on),
+    rift = recipe.package.startswith("rift.")
+    patches = [p for p in base.all_patches() if (p.category == "pcvr") == rift]
+    out = {}
+    for category in ("overport", "frame", "pcvr"):
+        out[category] = sorted(p.id for p in patches if p.category == category and not p.default_on and p.id in pats)
+        out[category + "_removed"] = sorted(p.id for p in patches
+                                            if p.category == category and p.default_on and p.id not in pats
+                                            and (defaults is None or p.id in defaults))
+    out.update({
         "adapter": {p.split(".", 1)[1]: v.get("value") for p, v in sorted(pats.items()) if p.startswith("adapter.")},
+        "device": sorted(p for p in pats if p.startswith("device.") and p not in DEVICE_FIELDS
+                         and not base.get(p).default_on),
         "device_files": sorted((pats.get("device.files") or {}).get("files", {})),
+        "foveation": dict(pats.get("device.foveation") or {}),
+        "lepton_env": dict((pats.get("device.lepton_env") or {}).get("env", {})),
+        "config_sync": dict((pats.get("device.config_sync") or {}).get("files", {})),
         "alt": sorted(recipe.alt_patches),
         "use_alt": recipe.use_alt,
         "status_unsupported": recipe.status == "unsupported",
-    }
+    })
+    return out
+
+
+def catalog_recipe(entry: catalog.CatalogEntry, defaults: set[str] | None = None, analysis=None) -> Recipe:
+    """Materialize catalog choices plus expected defaults; never import heuristic opt-in choices. With the game's
+    analysis, the engine's structural transforms apply as in suggest(): a 32-bit game goes through the bridge (its
+    entry relies on that instead of listing what the bridge replaces), a 2D app keeps only its non-VR patches."""
+    r = Recipe(entry.package, status=entry.status, as_is=entry.as_is,
+               alt_patches=list(entry.alt_overport), use_alt=entry.use_alt)
+    rift = entry.kind == "rift" or entry.package.startswith("rift.")
+    r.patches = {p.id: {} for p in base.all_patches() if p.default_on and (p.category == "pcvr") == rift
+                 and (defaults is None or p.id in defaults)}
+    for pid in (*entry.overport_extra, *entry.frame, *entry.device, *entry.pcvr):
+        r.patches[pid] = {}
+    for pid in (*entry.overport_remove, *entry.frame_remove, *entry.pcvr_remove):
+        r.patches.pop(pid, None)
+    for key, value in entry.adapter.items():
+        r.patches[f"adapter.{key}"] = {"value": value}
+    for pid, param, values in (("device.files", "files", entry.device_files),
+                               ("device.lepton_env", "env", entry.lepton_env),
+                               ("device.config_sync", "files", entry.config_sync)):
+        if values:
+            r.patches[pid] = {param: copy.deepcopy(values)}
+    if entry.foveation:
+        r.patches["device.foveation"] = {"mode": entry.foveation}
+        if entry.foveation_radius:
+            r.patches["device.foveation"]["radius"] = entry.foveation_radius
+        if entry.foveation_test:
+            r.patches["device.foveation"]["test"] = True
+    if analysis is not None:
+        if analysis.only_32bit and "frame.a32_bridge" not in entry.frame_remove:
+            r.patches.setdefault("frame.a32_bridge", {})
+            engine.a32_recipe(analysis, r, newly=False)
+        elif analysis.vr_kind == "none":
+            needed = {pid for pid in r.patches if not base.get(pid).needs_vr and pid != "frame.launcher"}
+            needed |= {"frame.launcher"} if analysis.has_info_category else set()
+            r.patches = {pid: v for pid, v in r.patches.items() if pid in needed}
+            r.alt_patches = []
+    return r
+
+
+def catalog_signature(entry: catalog.CatalogEntry, defaults: set[str] | None = None, analysis=None) -> dict:
+    return signature(catalog_recipe(entry, defaults, analysis), defaults)
+
+
+def engine_group(name: str) -> str:
+    return name if name in ("Unity", "Unreal") else "Other"
+
+
+def split_packages(engines: dict[str, str]) -> dict[str, str]:
+    """package -> dev/holdout; nearest 20% per engine (half rounded up), lowest sha256 first.
+
+    Only pass the intersection of catalog and analyses. This is stable under input reordering; adding packages can
+    change membership at a stratum's cutoff. Tiny strata may have no holdout. Unsupported entries participate too.
+    """
+    groups = defaultdict(list)
+    for pkg, name in engines.items():
+        groups[engine_group(name)].append(pkg)
+    result = {}
+    for packages in groups.values():
+        ordered = sorted(packages, key=lambda p: (hashlib.sha256(p.encode()).digest(), p))
+        n = (len(ordered) + 2) // 5
+        result.update((p, "holdout" if i < n else "dev") for i, p in enumerate(ordered))
+    return dict(sorted(result.items()))
+
+
+def _parts(sig: dict) -> tuple[dict, dict]:
+    headline, runtime = {}, {}
+    for key, value in sig.items():
+        if key == "adapter":
+            headline[key] = {k: v for k, v in value.items() if f"adapter.{k}" not in RUNTIME_ONLY}
+            runtime[key] = {k: v for k, v in value.items() if f"adapter.{k}" in RUNTIME_ONLY}
+        else:
+            (runtime if key in RUNTIME_ONLY else headline)[key] = value
+    return headline, runtime
+
+
+def _decisions(sig: dict) -> dict:
+    """Patch/action identities -> values. Status is a field metric, not a patch."""
+    out = {}
+    for field, value in sig.items():
+        if field == "status_unsupported":
+            continue
+        if field in ("adapter", "lepton_env", "config_sync"):
+            prefix = "adapter" if field == "adapter" else f"device.{field}"
+            out.update((f"{prefix}.{k}", v) for k, v in value.items())
+        elif field == "foveation":
+            if value:
+                out["device.foveation"] = value
+        elif field == "use_alt":
+            if value:
+                out["use_alt"] = value
+        else:
+            prefix = ("remove:" if field.endswith("_removed") else "alt:" if field == "alt"
+                      else "device.files:" if field == "device_files" else "")
+            out.update((prefix + p, True) for p in value)
+    return out
+
+
+def _rates(tp: int, fp: int, fn: int) -> dict:
+    # No positive decisions yields 0, rather than inflating scores for empty recipes/strata.
+    return {"tp": tp, "fp": fp, "fn": fn, "precision": tp / (tp + fp) if tp + fp else 0.0,
+            "recall": tp / (tp + fn) if tp + fn else 0.0,
+            "f1": 2 * tp / (2 * tp + fp + fn) if 2 * tp + fp + fn else 0.0}
+
+
+def score_signatures(want: dict, got: dict, gaps: frozenset | set = frozenset()) -> dict:
+    """Compare one pair of signatures; value mismatches count as both missing and extra. `gaps` = decisions the
+    catalog entry couldn't have made (the patch didn't exist when it was verified): left out of both sides."""
+    result = {}
+    for scope, w, g in zip(("headline", "runtime"), _parts(want), _parts(got), strict=True):
+        wanted, actual = _decisions(w), _decisions(g)
+        gap = {k: actual.pop(k) for k in list(actual) if k in gaps and k not in wanted}
+        missing = {k: v for k, v in wanted.items() if k not in actual or actual[k] != v}
+        extra = {k: v for k, v in actual.items() if k not in wanted or wanted[k] != v}
+        per_patch = {k: _rates(int(k in wanted and k in actual and wanted[k] == actual[k]),
+                               int(k in extra), int(k in missing)) for k in sorted(wanted.keys() | actual.keys())}
+        fields = {k: w.get(k) == g.get(k) for k in sorted(w.keys() | g.keys())}
+        exact = all(fields.values()) if not gap else not missing and not extra
+        result[scope] = {"exact": exact, "fields": fields, "missing": missing, "extra": extra,
+                         "per_patch": per_patch, "gap": gap}
+    return result
+
+
+def _aggregate(rows: list[dict], scope: str, score_key: str = "score") -> dict:
+    fields = defaultdict(lambda: {"matched": 0, "total": 0})
+    counts = defaultdict(Counter)
+    gaps = Counter()
+    for row in rows:
+        s = row[score_key][scope]
+        gaps.update(s.get("gap") or {})
+        for key, match in s["fields"].items():
+            fields[key]["matched"] += int(match)
+            fields[key]["total"] += 1
+        for key, stats in s["per_patch"].items():
+            counts[key].update({k: stats[k] for k in ("tp", "fp", "fn")})
+    per_patch = {k: _rates(**v) for k, v in sorted(counts.items())}
+    total = Counter()
+    for stats in counts.values():
+        total.update(stats)
+    exact = sum(row[score_key][scope]["exact"] for row in rows)
+    return {"games": len(rows), "exact": exact, "exact_ratio": exact / len(rows) if rows else 0.0,
+            "fields": dict(sorted(fields.items())), "per_patch": per_patch, "gaps": dict(sorted(gaps.items())),
+            "patches": _rates(*(total[k] for k in ("tp", "fp", "fn")))}
+
+
+def score_records(records: list[dict]) -> dict:
+    """Score {package, engine, status, converted, want, got, ...} records; no filesystem needed."""
+    rows = [{**r, "score": score_signatures(r["want"], r["got"]),
+             "score_dated": score_signatures(r["want"], r["got"], set(r.get("gaps") or ()))}
+            for r in sorted(records, key=lambda r: r["package"])]
+
+    def metrics(group):
+        supported = [r for r in group if r["status"] != "unsupported"]
+        unsupported = [r for r in group if r["status"] == "unsupported"]
+        return {"total_games": len(group), "converted_games": sum(bool(r.get("converted")) for r in group),
+                "excluded_unsupported": len(unsupported),
+                "headline": _aggregate(supported, "headline"), "runtime": _aggregate(supported, "runtime"),
+                "headline_dated": _aggregate(supported, "headline", "score_dated"),
+                "unsupported": {scope: _aggregate(unsupported, scope) for scope in ("headline", "runtime")}}
+
+    return {**metrics(rows), "by_engine": {name: metrics([r for r in rows if engine_group(r["engine"]) == name])
+                                           for name in ("Unity", "Unreal", "Other")}, "games": rows}
+
+
+def read_snapshot(path: Path, attempts: int = 8) -> dict:
+    """Copy before parsing, even on retries; never write to the source or read APK/data paths in it."""
+    with tempfile.TemporaryDirectory(prefix="frameport-eval-") as tmp:
+        dest = Path(tmp) / "snapshot.json"
+        for attempt in range(attempts):
+            shutil.copyfile(path, dest)
+            try:
+                return json.loads(dest.read_text(encoding="utf-8"))
+            except json.JSONDecodeError:
+                if attempt == attempts - 1:
+                    raise
+                time.sleep(0.25 * (attempt + 1))
+    raise AssertionError("unreachable")
+
+
+def read_dumps(path: Path) -> dict:
+    records = {}
+    for g in quest_dump.scan(path):
+        a = analyze(g.apk, data_bytes=g.data_bytes())
+        row = {"analysis": a.to_dict(), "converted": a.is_overport_output, "name": g.name}
+        if a.package not in records or records[a.package]["converted"] and not row["converted"]:
+            records[a.package] = row
+    return records
+
+
+def gap_decisions(entry, intro_dates: dict | None) -> set[str]:
+    """Decision keys for patches introduced after the entry was verified: the catalog can't be wrong about them."""
+    verified = str((entry.verified or {}).get("date") or "")
+    if not intro_dates or not verified:
+        return set()
+    return {(f"alt:{pid}" if pid.startswith("patch_remove_unreal") else pid)
+            for pid, date in intro_dates.items() if date and date > verified}
+
+
+def inherited_decisions() -> set[str]:
+    """Heuristic picks a catalog recipe keeps even when the entry doesn't name them (engine.suggest: adapter
+    settings and device patches other than catalog.TOGGLED_DEVICE): the verified build had them, so an "extra" one
+    isn't a detection error. A catalog value that differs still counts (the key is in both sides then)."""
+    return {p.id for p in base.all_patches()
+            if p.category == "adapter" or (p.category == "device" and p.id not in catalog.TOGGLED_DEVICE)}
+
+
+def evaluate(snapshot: dict, entries: dict, split: str = "dev", intro_dates: dict | None = None) -> dict:
+    analyses = {pkg: analysis_from_dict(v["analysis"]) for pkg, v in snapshot.items() if pkg in entries}
+    assignments = split_packages({pkg: a.engine for pkg, a in analyses.items()})
+    records = []
+    for pkg, a in sorted(analyses.items()):
+        if split != "all" and assignments[pkg] != split:
+            continue
+        entry = entries[pkg]
+        defaults = expected_defaults(a)
+        records.append({"package": pkg, "title": entry.title, "engine": a.engine, "status": entry.status,
+                        "converted": bool(snapshot[pkg].get("converted", a.is_overport_output)),
+                        "analysis_version": snapshot[pkg].get("analysis_version"), "split": assignments[pkg],
+                        "want": catalog_signature(entry, defaults, a),
+                        "got": signature(engine.suggest(a, use_catalog=False), defaults),
+                        "gaps": sorted(gap_decisions(entry, intro_dates) | inherited_decisions())})
+    return {"schema_version": 1, "split": split, "runtime_only": RUNTIME_ONLY, "snapshot_packages": len(snapshot),
+            "default_policy": "per-analysis default detection, then explicit catalog removals",
+            "catalog_packages": len(entries), "matched_packages": len(analyses),
+            "split_assignments": assignments, **score_records(records)}
+
+
+def fit_candidates(analysis, entry: catalog.CatalogEntry) -> list[dict]:
+    """Use suggest's final heuristic pass, including platform filtering, non-VR and a32 normalization.
+
+    Requiring a recommended detect() as well excludes dependencies and synthetic recipe additions. No engine
+    refactor is needed: suggest() remains untouched, and catalog data is used only for the candidate exclusion and
+    the hypothetical warning recipe.
+    """
+    suggested = engine.suggest(analysis, use_catalog=False)
+    chosen = catalog_recipe(entry, expected_defaults(analysis), analysis)
+    listed = set(chosen.patches) | set(entry.alt_overport)
+    removed = set(entry.overport_remove) | set(entry.frame_remove) | set(entry.pcvr_remove)
+    result = []
+    for pid, params in sorted(suggested.patches.items()):
+        p = base.get(pid)
+        if (p.default_on or pid in listed or pid in removed or not base.for_game(p, analysis)
+                or not p.applies(analysis)):
+            continue
+        s = p.detect(analysis)
+        if not s or not s.recommended:
+            continue
+        trial = copy.deepcopy(chosen)
+        trial.patches[pid] = dict(params)
+        result.append({"patch": pid, "category": p.category, "experimental": p.experimental,
+                       "strict": bool(getattr(p, "strict", False)), "reason": s.reason,
+                       "warnings": engine.warnings(trial, analysis)})
+    return result
+
+
+def patch_intro_dates(cache_path: Path) -> dict:
+    """One local git pickaxe per uncached patch (including missing-history results); never use the network."""
+    try:
+        cached = json.loads(cache_path.read_text(encoding="utf-8"))
+    except (OSError, ValueError):
+        cached = {}
+    changed = False
+    for p in base.all_patches():
+        if p.id in cached:
+            continue
+        needle = p.id.split(".", 1)[1] if p.category == "adapter" else p.id
+        paths = ["src/frameport/patches/settings.py"] if p.category == "adapter" else ["src/frameport"]
+        log = subprocess.run(["git", "log", f'-S"{needle}"', "--reverse", "--format=%ad", "--date=short", "--",
+                              *paths], cwd=REPO, text=True, capture_output=True, check=True)
+        dates = log.stdout.splitlines()
+        cached[p.id] = dates[0] if dates else None
+        changed = True
+    if changed:
+        cache_path.write_text(json.dumps(cached, indent=2, sort_keys=True) + "\n", encoding="utf-8")
+    return cached
+
+
+def entry_date(entry) -> str:
+    """When the recipe was verified; `updated` only when there is no verified date (it also moves for unrelated
+    edits, e.g. a note, so it says nothing about which patches the verifier could pick)."""
+    return str(entry.verified.get("date") or "") or str(entry.updated or "")
+
+
+def fit_report(snapshot: dict, entries: dict, intro_dates: dict) -> dict:
+    games, metadata = [], []
+    totals = Counter()
+    for pkg, entry in sorted(entries.items()):
+        common = {"package": pkg, "title": entry.title, "engine": entry.engine, "xr": entry.xr,
+                  "status": entry.status, "verified_date": str(entry.verified.get("date") or ""),
+                  "updated": str(entry.updated or ""), "entry_date": entry_date(entry)}
+        if pkg not in snapshot:
+            metadata.append(common)
+            continue
+        candidates = fit_candidates(analysis_from_dict(snapshot[pkg]["analysis"]), entry)
+        for c in candidates:
+            intro = intro_dates.get(c["patch"])
+            c.update(intro_date=intro,
+                     newer_patch=bool(intro and common["entry_date"] and intro > common["entry_date"]))
+            totals[c["patch"]] += 1
+        games.append({**common, "converted": bool(snapshot[pkg].get("converted")), "candidates": candidates})
+    return {"games": games, "metadata_only": metadata, "totals_by_patch": dict(sorted(totals.items())),
+            "candidate_count": sum(totals.values()), "games_with_candidates": sum(bool(g["candidates"]) for g in games)}
+
+
+def _md(value) -> str:
+    return str(value if value is not None else "unknown").replace("|", "\\|").replace("\n", " ")
+
+
+def render_fit_report(report: dict) -> str:
+    lines = ["# New patch fit", "", f"{len(report['games'])} catalog games with analyses; "
+             f"{report['games_with_candidates']} with candidates; {report['candidate_count']} candidate additions. "
+             f"{len(report['metadata_only'])} metadata-only entries.", "",
+             "Candidates are opt-in recommended patches absent from the entry, after the heuristic engine's platform "
+             "and a32 rules. A listed adapter key is excluded even if its value differs. Warnings are for the full "
+             "catalog recipe plus that one patch; they can include pre-existing conflicts. Unknown history/date "
+             "does not imply a newer patch. Default-on patches are not candidates. All splits and statuses included.",
+             "", "## Totals by patch", "", "| Patch | Games |", "| --- | ---: |"]
+    lines.extend(f"| {pid} | {n} |" for pid, n in
+                 sorted(report["totals_by_patch"].items(), key=lambda kv: (-kv[1], kv[0])))
+    for game in report["games"]:
+        lines += ["", f"## {_md(game['title'])} (`{game['package']}`)", "",
+                  f"{game['engine']} / {game['xr']}; status: {game['status']}; "
+                  f"entry date: {game['entry_date'] or 'unknown'}; converted: {game['converted']}.", ""]
+        if not game["candidates"]:
+            lines += ["No candidates."]
+            continue
+        lines += ["| Patch | Category | Experimental | Strict | Reason | Introduced | Newer patch | Warnings |",
+                  "| --- | --- | --- | --- | --- | --- | --- | --- |"]
+        for c in game["candidates"]:
+            cells = [c["patch"], c["category"], c["experimental"], c["strict"], c["reason"], c["intro_date"],
+                     c["newer_patch"], "; ".join(c["warnings"]) or "none"]
+            lines.append("| " + " | ".join(_md(v) for v in cells) + " |")
+    lines += ["", "## Catalog entries without analyses (metadata only)", "",
+              "| Package | Title | Engine | XR | Status | Verified | Updated | Entry date |",
+              "| --- | --- | --- | --- | --- | --- | --- | --- |"]
+    for g in report["metadata_only"]:
+        lines.append("| " + " | ".join(_md(g[k]) for k in
+                     ("package", "title", "engine", "xr", "status", "verified_date", "updated", "entry_date")) + " |")
+    return "\n".join(lines) + "\n"
+
+
+def print_metrics(report: dict, verbose: bool = False) -> None:
+    print(f"Split: {report['split']}; catalog/analysis intersection: {report['matched_packages']}; "
+          f"converted: {report['converted_games']}; unsupported excluded: {report['excluded_unsupported']}")
+    print(f"{'Engine':10} {'Games':>5} {'Exact':>9} {'Fields':>11} {'Prec':>7} {'Recall':>7} {'F1':>7} {'Runtime':>9}")
+    for name, metrics in [("All", report), *report["by_engine"].items()]:
+        h, r = metrics["headline"], metrics["runtime"]
+        fields = h["fields"].values()
+        f = f"{sum(x['matched'] for x in fields)}/{sum(x['total'] for x in h['fields'].values())}"
+        p = h["patches"]
+        print(f"{name:10} {h['games']:5} {h['exact']:4}/{h['games']:<4} {f:>11} "
+              f"{p['precision']:7.1%} {p['recall']:7.1%} {p['f1']:7.1%} {r['exact']:4}/{r['games']:<4}")
+    print("\nDated (extras the entry couldn't have listed - patch newer than its verified date - or that catalog "
+          "recipes inherit anyway - adapter/device picks - are not errors):")
+    for name, metrics in [("All", report), *report["by_engine"].items()]:
+        d = metrics["headline_dated"]
+        p = d["patches"]
+        print(f"{name:10} {d['games']:5} {d['exact']:4}/{d['games']:<4} {'':>11} "
+              f"{p['precision']:7.1%} {p['recall']:7.1%} {p['f1']:7.1%}")
+    gaps = report["headline_dated"]["gaps"]
+    if gaps:
+        print("Catalog gaps by patch: " + ", ".join(f"{k}={v}" for k, v in sorted(gaps.items(), key=lambda kv: -kv[1])))
+    for name, metrics in [("All", report), *report["by_engine"].items()]:
+        print(f"\n{name}: field matches (headline / runtime)")
+        for scope in ("headline", "runtime"):
+            matches = metrics[scope]["fields"].items()
+            print(scope + ": " + ", ".join(f"{k}={v['matched']}/{v['total']}" for k, v in matches))
+        print(f"{'Scope / patch':55} {'TP':>5} {'FP':>5} {'FN':>5} {'Prec':>7} {'Recall':>7} {'F1':>7}")
+        for scope in ("headline", "runtime"):
+            for pid, p in metrics[scope]["per_patch"].items():
+                print(f"{scope + ' / ' + pid:55} {p['tp']:5} {p['fp']:5} {p['fn']:5} "
+                      f"{p['precision']:7.1%} {p['recall']:7.1%} {p['f1']:7.1%}")
+    u = report["unsupported"]["headline"]
+    print(f"\nUnsupported (excluded): {u['exact']}/{u['games']} exact; patch F1={u['patches']['f1']:.1%}")
+    for game in report["games"]:
+        s = game["score"]
+        if verbose or not s["headline"]["exact"] or not s["runtime"]["exact"] or game["status"] == "unsupported":
+            print(f"\n{game['package']} [{game['engine']}, {game['status']}, {game['split']}"
+                  + (", converted" if game["converted"] else "") + "]")
+            for scope in ("headline", "runtime"):
+                for diff in ("missing", "extra"):
+                    if s[scope][diff]:
+                        print(f"  {scope} {diff}: {json.dumps(s[scope][diff], sort_keys=True)}")
+                if not s[scope]["fields"].get("status_unsupported", True):
+                    print(f"  {scope} status_unsupported: catalog={game['want']['status_unsupported']} "
+                          f"heuristics={game['got']['status_unsupported']}")
 
 
 def main() -> int:
-    ap = argparse.ArgumentParser()
-    ap.add_argument("dumps", type=Path)
+    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument("dumps", type=Path, nargs="?")
+    ap.add_argument("--snapshot", type=Path)
+    ap.add_argument("--split", choices=("dev", "holdout", "all"), default="dev")
+    ap.add_argument("--json", type=Path, dest="json_out")
     ap.add_argument("--verbose", action="store_true")
+    ap.add_argument("--fit-report", action="store_true")
+    ap.add_argument("--out", type=Path, help="Markdown destination for --fit-report (otherwise stdout)")
     args = ap.parse_args()
-    games = quest_dump.scan(args.dumps)
-    total = matched_fields = games_ok = 0
-    misses = []
-    for g in games:
-        a = analyze(g.apk, data_bytes=g.data_bytes())
-        if not catalog.lookup(a.package):
-            continue
-        want = signature(engine.suggest(a, use_catalog=True))
-        want["status_unsupported"] = (catalog.lookup(a.package).status == "unsupported"
-                                      and "32-bit" in catalog.lookup(a.package).notes)
-        got = signature(engine.suggest(a, use_catalog=False))
-        diffs = {k: (want[k], got[k]) for k in want if want[k] != got[k]}
-        total += len(want)
-        matched_fields += len(want) - len(diffs)
-        games_ok += not diffs
-        if diffs:
-            misses.append((a.package, diffs))
-        if args.verbose or diffs:
-            print(f"{'OK  ' if not diffs else 'DIFF'} {a.package}")
-            for k, (w, h) in diffs.items():
-                print(f"       {k}: catalog={w}  heuristics={h}")
-    n = games_ok + len(misses)
-    print(f"\n{games_ok}/{n} games reproduced exactly; {matched_fields}/{total} recipe fields match")
-    return 0 if not misses else 1
+    if not args.snapshot and not args.dumps:
+        ap.error("provide a dumps folder or --snapshot")
+    if args.out and not args.fit_report:
+        ap.error("--out requires --fit-report")
+    snapshot = read_snapshot(args.snapshot) if args.snapshot else read_dumps(args.dumps)
+    # Explicitly bundled entries: remote caches and personal recipes must not affect this benchmark.
+    entries = catalog._load_dir(REPO / "catalog/games", "bundled")
+    if args.fit_report:
+        cache = (args.snapshot.parent if args.snapshot else args.dumps) / "patch-intro-dates.json"
+        report = fit_report(snapshot, entries, patch_intro_dates(cache))
+        markdown = render_fit_report(report)
+        if args.out:
+            args.out.write_text(markdown, encoding="utf-8")
+            print(f"{report['candidate_count']} candidates in {report['games_with_candidates']} games -> {args.out}")
+        else:
+            print(markdown, end="")
+    else:
+        cache = (args.snapshot.parent if args.snapshot else args.dumps) / "patch-intro-dates.json"
+        # dated scoring reads the cache --fit-report writes; a plain run never shells out for dates
+        dates = json.loads(cache.read_text(encoding="utf-8")) if cache.is_file() else None
+        report = evaluate(snapshot, entries, args.split, dates)
+        print_metrics(report, args.verbose)
+    if args.json_out:
+        args.json_out.write_text(json.dumps(report, indent=2, sort_keys=True, default=str) + "\n", encoding="utf-8")
+    # Retain the legacy evaluator's nonzero exit for recipe differences; JSON is still written.
+    return 0 if args.fit_report or report["headline"]["games"] == report["headline"]["exact"] else 1
 
 
 if __name__ == "__main__":

scripts/heuristics_snapshot.py

@@ -0,0 +1,79 @@
+#!/usr/bin/env python3
+"""Analyse every Quest dump below the given folders once and save the analyses as JSON, so heuristics evaluations
+(`eval_heuristics.py --snapshot`) don't re-read gigabytes of APKs each run.
+
+    python scripts/heuristics_snapshot.py <out.json> <dumps folder> [<dumps folder> ...]
+
+An original APK beats an already converted one (OVRPort output, e.g. PATCHED/ copies): converted inputs change what
+the heuristics suggest, so they are only kept when no original exists and are marked `converted`. The output holds
+local paths: keep it out of git.
+"""
+from __future__ import annotations
+
+import argparse
+import dataclasses
+import json
+import sys
+import time
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[1] / "src"))
+
+from frameport.analysis.detect import ANALYSIS_VERSION, analyze  # noqa: E402
+from frameport.sources import quest_dump  # noqa: E402
+
+
+def _skip(game, roots: list[Path]) -> bool:
+    """Folders starting with "_" (known-good backups, test builds, keys) are not dumps."""
+    for root in roots:
+        try:
+            rel = game.apk.relative_to(root)
+        except ValueError:
+            continue
+        return any(part.startswith("_") for part in rel.parts[:-1])
+    return False
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("out", type=Path)
+    ap.add_argument("roots", type=Path, nargs="+")
+    args = ap.parse_args()
+    roots = [r.resolve() for r in args.roots]
+    old = json.loads(args.out.read_text()) if args.out.is_file() else {}
+    snap: dict[str, dict] = {}
+    for root in roots:
+        for g in quest_dump.scan(root):
+            if _skip(g, roots):
+                continue
+            key = str(g.apk)
+            cached = next((v for v in old.values() if v.get("apk") == key
+                           and v.get("analysis_version") == ANALYSIS_VERSION
+                           and v.get("apk_size") == g.apk.stat().st_size), None)
+            t = time.time()
+            if cached:
+                entry = cached
+            else:
+                try:
+                    data_bytes = g.data_bytes()
+                    a = analyze(g.apk, data_bytes=data_bytes)
+                except Exception as e:  # noqa: BLE001 - one unreadable dump mustn't stop the snapshot
+                    print(f"FAIL {g.apk}: {e}", file=sys.stderr, flush=True)
+                    continue
+                entry = {"analysis": dataclasses.asdict(a), "data_bytes": data_bytes, "apk": key,
+                         "apk_size": g.apk.stat().st_size, "name": g.name, "root": str(root),
+                         "converted": bool(a.is_overport_output), "analysis_version": ANALYSIS_VERSION}
+            pkg = entry["analysis"]["package"]
+            have = snap.get(pkg)
+            # originals beat converted copies; otherwise the first root given wins
+            if have is None or (have["converted"] and not entry["converted"]):
+                snap[pkg] = entry
+            print(f"{'cache' if cached else f'{time.time() - t:5.1f}s'} {pkg}"
+                  f"{' (converted)' if entry['converted'] else ''}", flush=True)
+            args.out.write_text(json.dumps(snap, indent=1, default=str))  # partial results survive a crash
+    print(f"{len(snap)} packages ({sum(v['converted'] for v in snap.values())} converted only) -> {args.out}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

src/frameport/patches/frame/unity_oculus_check.py

@@ -112,6 +112,7 @@ class UnityOculusCheck(Patch):
         return a.engine == "Unity" and "libOVRPlugin.so" in a.libs and bool((a.extra or {}).get("unity_oculus_check"))
 
     def detect(self, a):
+        # not narrowed by libOculusXRPlugin.so: a verified game ships it and still needs this fix
         if self.applies(a):
             return Suggestion(True, "Unity with built-in Oculus support: it checks for Meta's system apps before "
                                     "starting VR, else it runs as a 2D app (for example Accounting+, BattleSisters).")

src/frameport/patches/frame/unity_runtime.py

@@ -12,7 +12,7 @@ frame.unity_text_input).
 from __future__ import annotations
 
 from ..base import Suggestion, register
-from .unity_text_input import RET_FALSE, Il2cppReturnPatch
+from .unity_text_input import RET_FALSE, Il2cppReturnPatch, can_rewrite_il2cpp
 
 RET_ZERO = RET_FALSE  # mov w0, #0 ; ret
 
@@ -40,7 +40,8 @@ class UnityRuntimeMsaa(Il2cppReturnPatch):
         return a.engine == "Unity" and "libil2cpp.so" in a.libs and not a.only_32bit
 
     def detect(self, a):
-        if self.applies(a) and (a.extra or {}).get("ovr_runtime_msaa") and "GLES" in a.graphics:
+        if (self.applies(a) and can_rewrite_il2cpp(a)
+                and (a.extra or {}).get("ovr_runtime_msaa") and "GLES" in a.graphics):
             return Suggestion(True, "GLES Unity game whose OVRManager turns 4x MSAA on at runtime: multisampled "
                                     "render-to-texture can hang the Frame's GPU.")
         return None

src/frameport/patches/frame/unity_text_input.py

@@ -85,6 +85,19 @@ def patch_field_loads(lib: bytes, method: tuple[int, int], field_offset: int, va
 ALL_TARGETS: dict[str, list[str]] = {}  # every IL2CPP patch's methods: one Cpp2IL run (and cache entry) serves all
 
 
+def can_rewrite_il2cpp(a) -> bool:
+    """Automatic return patches need AArch64 code and metadata Cpp2IL can read.
+
+    Unity 2017's metadata is unreadable by our Cpp2IL workflow (AGENTS.md, Accounting+). Unknown versions keep the
+    existing fallback: apply() reads the version from the APK. This does not restrict manually selected patches.
+    """
+    try:
+        major = int(str((a.extra or {}).get("unity_version") or "0").split(".")[0])
+    except ValueError:
+        major = 0
+    return "arm64-v8a" in a.abis and not 0 < major < 2018
+
+
 class Il2cppReturnPatch(Patch):
     """Make IL2CPP methods return a constant (targets), found per game with Cpp2IL (analysis/il2cpp.py)."""
     targets: dict[str, dict[str, bytes]] = {}
@@ -164,7 +177,7 @@ class UnityTextInput(Il2cppReturnPatch):
         return a.engine == "Unity" and "libil2cpp.so" in a.libs and bool((a.extra or {}).get("text_fields"))
 
     def detect(self, a):
-        if self.applies(a):
+        if self.applies(a) and can_rewrite_il2cpp(a):
             kinds = " and ".join((a.extra or {}).get("text_fields"))
             return Suggestion(True, f"Unity app with text fields ({kinds}): they close at once on the Frame because "
                                     "it has no system keyboard.")

src/frameport/patches/frame/unity_tls_key_pool.py

@@ -40,6 +40,13 @@ class UnityTlsKeyPool(Patch):
     def detect(self, a):
         heavy = heavy_native_libs(a)
         if self.applies(a) and _unity_major(a) >= 6000 and heavy:
+            # Firebase's Firestore and Gecko both consume keys alongside Unity 6 and the Frame's graphics/XR
+            # runtime. This is the documented exhaustion case (AGENTS.md, #198), rather than just any Firebase
+            # library: keep a lone SDK as an optional diagnostic fix below.
+            if "libFirebaseCppFirestore.so" in heavy and "libxul.so" in heavy:
+                return Suggestion(True, "Unity 6 with Firebase Firestore and Gecko: these native runtimes can "
+                                        "exhaust Android's 128 thread keys alongside the Frame's graphics and XR "
+                                        "libraries. Pool Unity's keys to leave room for them.")
             return Suggestion(False, f"Unity 6 with {', '.join(heavy)}: turn on if the game aborts at start with "
                                      "\"cannot create thread specific key\" (for example Wander).")
         return None

src/frameport/patches/settings.py

@@ -486,6 +486,11 @@ class AdapterSetting(Patch):
                                     "Tale, BattleSisters): vibrations are converted before they reach it.",
                               {"value": 1})
         if (self.key == "pose_time_fix" and a.engine == "Unity" and "libOVRPlugin.so" in a.libs
+                # Absence of the Oculus XR Plugin isn't proof of Unity's built-in VR: Unity OpenXR games can
+                # bundle OVRPlugin too. The system-app check in libunity is positive evidence of built-in Oculus
+                # (analysis/detect.py, frame/unity_oculus_check.py). Older analyses lack this signal; converted
+                # inputs may have had the check rewritten to "android", so keep the legacy fallback for those.
+                and (a.extra.get("unity_oculus_check", True) or a.is_overport_output)
                 and not a.extra.get("oculus_xr_plugin") and "libOculusXRPlugin.so" not in a.libs):
             return Suggestion(True, "Unity's built-in Oculus support reads controller poses through OVRPlugin at its "
                                     "monotonic \"now\", seconds in the past on the Frame: they are located at the "

tests/test_eval_heuristics.py

@@ -0,0 +1,326 @@
+"""Synthetic evaluator tests: no APKs, snapshots, git history, or network needed."""
+import importlib.util
+import json
+from pathlib import Path
+
+import pytest
+
+from frameport.core.library import analysis_from_dict
+from frameport.patches import base
+from frameport.recommend import catalog, engine
+
+script = Path(__file__).parents[1] / "scripts/eval_heuristics.py"
+spec = importlib.util.spec_from_file_location("eval_heuristics", script)
+ev = importlib.util.module_from_spec(spec)
+spec.loader.exec_module(ev)
+
+
+def _analysis(**kwargs):
+    fields = dict(package="com.example.game", engine="Unity", xr="OpenXR", abis=["arm64-v8a"],
+                  libs=[], extra={}, graphics="Vulkan")
+    fields.update(kwargs)
+    return analysis_from_dict(fields)
+
+
+def _entry(**kwargs):
+    return catalog.CatalogEntry(**(dict(package="com.example.game", title="Example", status="works") | kwargs))
+
+
+def _patch(monkeypatch, pid, *, category="frame", default=False, applies=True, recommended=True,
+           a32=None, requires=(), conflicts=()):
+    base.load_all()
+    p = base.Patch()
+    p.id, p.title, p.category, p.default_on = pid, pid, category, default
+    p.a32, p.requires, p.conflicts = a32, requires, conflicts
+    p.experimental, p.strict = True, True
+    p.applies = lambda _: applies
+    params = {"value": 3} if category == "adapter" else {}
+    p.detect = lambda _: base.Suggestion(recommended, f"reason for {pid}", params)
+    monkeypatch.setitem(base.REGISTRY, pid, p)
+    return p
+
+
+def test_catalog_signature_is_catalog_only(monkeypatch):
+    def forbidden(*args, **kwargs):
+        pytest.fail("catalog signature must not call suggest")
+
+    monkeypatch.setattr(engine, "suggest", forbidden)
+    a = _analysis()
+    defaults = ev.expected_defaults(a)
+    entry = _entry(overport_extra=["patch_remove_unreal_force_quit"], overport_remove=["patch_copy_libraries"],
+                   frame=["frame.unity_text_input"], frame_remove=["frame.adapter"],
+                   adapter={"scene_emul": 0, "scale": 1.5}, device=["device.text_input_window"],
+                   device_files={"user.cfg": "r_vsync=0"}, alt_overport=["patch_fix_ue4_crash"], use_alt=True,
+                   foveation="fixed", foveation_radius=0.5, foveation_test=True,
+                   lepton_env={"X": "1"}, config_sync={"user.cfg": {"width": "w"}})
+    want = ev.catalog_signature(entry, defaults)
+    assert want["overport"] == ["patch_remove_unreal_force_quit"]
+    assert want["overport_removed"] == ["patch_copy_libraries"]
+    assert want["frame"] == ["frame.unity_text_input"]
+    assert want["frame_removed"] == ["frame.adapter"]
+    assert want["adapter"] == {"scene_emul": 0, "scale": 1.5}
+    assert want["device"] == ["device.text_input_window"]
+    assert want["device_files"] == ["user.cfg"]
+    assert want["alt"] == ["patch_fix_ue4_crash"]
+    assert want["use_alt"] is True
+    assert want["lepton_env"] == {"X": "1"}
+    assert want["config_sync"] == {"user.cfg": {"width": "w"}}
+    assert want["foveation"] == {"mode": "fixed", "radius": 0.5, "test": True}
+    assert not want["status_unsupported"]
+
+
+def test_mask_conditional_defaults_and_score_explicit_removals():
+    a = _analysis()
+    defaults = ev.expected_defaults(a)
+    assert "frame.ltw_depth" not in defaults
+    assert "frame.vk_sanitize" not in defaults
+    assert "frame.adapter" in defaults
+    r = ev.catalog_recipe(_entry(), defaults)
+    assert ev.signature(r, defaults)["frame_removed"] == []
+    removed = ev.catalog_signature(_entry(frame_remove=["frame.adapter", "frame.ltw_depth"]), defaults)
+    assert removed["frame_removed"] == ["frame.adapter"]
+    scored = ev.score_signatures(removed, ev.signature(r, defaults))["headline"]
+    assert scored["missing"] == {"remove:frame.adapter": True}
+    assert not scored["exact"]
+    r.patches.pop("frame.adapter")
+    dropped = ev.score_signatures(ev.catalog_signature(_entry(), defaults), ev.signature(r, defaults))["headline"]
+    assert dropped["extra"] == {"remove:frame.adapter": True}
+
+
+def test_catalog_adapter_does_not_inherit_heuristics():
+    a = _analysis(libs=["libOVRPlugin.so"])
+    got = engine.suggest(a, use_catalog=False)
+    assert "adapter.haptic_fix" in got.patches
+    assert ev.catalog_signature(_entry(), ev.expected_defaults(a))["adapter"] == {}
+
+
+def test_scorer_values_counts_and_exact():
+    want = ev.catalog_signature(_entry(frame=["frame.unity_text_input"], adapter={"scene_emul": 1}))
+    got = ev.catalog_signature(_entry(frame=["frame.unity_text_input", "frame.hw_video_decode"],
+                                      adapter={"scene_emul": 2}))
+    score = ev.score_signatures(want, got)["headline"]
+    assert not score["exact"]
+    assert score["per_patch"]["frame.unity_text_input"] == ev._rates(1, 0, 0)
+    assert score["per_patch"]["frame.hw_video_decode"] == ev._rates(0, 1, 0)
+    assert score["per_patch"]["adapter.scene_emul"] == ev._rates(0, 1, 1)
+    assert score["missing"] == {"adapter.scene_emul": 1}
+    assert score["extra"] == {"adapter.scene_emul": 2, "frame.hw_video_decode": True}
+    same = ev.score_signatures(want, want)
+    assert same["headline"]["exact"] and same["runtime"]["exact"]
+    assert same["headline"]["per_patch"]["adapter.scene_emul"] == ev._rates(1, 0, 0)
+
+
+def test_runtime_separate_from_headline():
+    want = ev.catalog_signature(_entry(adapter={"scale": 1.5, "sync_guard": 1}, use_alt=True,
+                                       lepton_env={"X": "Y"}, config_sync={"foo": {"w": "w"}}))
+    got = ev.catalog_signature(_entry())
+    score = ev.score_signatures(want, got)
+    assert score["headline"]["exact"]
+    assert not score["headline"]["missing"]
+    assert not score["runtime"]["exact"]
+    assert score["runtime"]["per_patch"]["adapter.scale"] == ev._rates(0, 0, 1)
+    assert score["runtime"]["missing"]["use_alt"]
+
+
+def test_aggregate_excludes_unsupported_and_breaks_down_engines():
+    want = ev.catalog_signature(_entry(frame=["frame.unity_text_input"]))
+    empty = ev.catalog_signature(_entry())
+    rows = [dict(package="com.a", status="works", engine="Unity", converted=True, want=want, got=want),
+            dict(package="com.b", status="issues", engine="Unreal", want=want, got=empty),
+            dict(package="com.c", status="unsupported", engine="Other", want=want, got=empty)]
+    result = ev.score_records(rows)
+    assert result["headline"]["games"] == 2
+    assert result["headline"]["exact"] == 1
+    assert result["headline"]["exact_ratio"] == 0.5
+    assert result["headline"]["patches"] == ev._rates(1, 0, 1)
+    assert result["headline"]["fields"]["frame"] == {"matched": 1, "total": 2}
+    assert result["runtime"]["exact"] == 2
+    assert result["excluded_unsupported"] == 1
+    assert result["unsupported"]["headline"]["patches"] == ev._rates(0, 0, 1)
+    assert result["by_engine"]["Unreal"]["headline"]["patches"] == ev._rates(0, 0, 1)
+    assert result["converted_games"] == 1
+    assert len(result["games"]) == 3
+
+
+def test_split_deterministic_stratified_and_disjoint():
+    engines = {f"com.example.{name}{i}": name for name in ("Unity", "Unreal", "Other") for i in range(10)}
+    assignments = ev.split_packages(engines)
+    assert assignments == ev.split_packages(dict(reversed(list(engines.items()))))
+    for name in ("Unity", "Unreal", "Other"):
+        mine = [pkg for pkg in engines if engines[pkg] == name]
+        ranked = sorted(mine, key=lambda p: ev.hashlib.sha256(p.encode()).digest())
+        assert {p for p in mine if assignments[p] == "holdout"} == set(ranked[:2])
+    dev = {p for p, s in assignments.items() if s == "dev"}
+    holdout = {p for p, s in assignments.items() if s == "holdout"}
+    assert not dev & holdout
+    assert dev | holdout == engines.keys()
+    assert ev.split_packages({}) == {}
+    assert ev.split_packages({"com.one": "CryEngine"}) == {"com.one": "dev"}
+
+
+def test_fit_selects_only_recommended_opt_ins_not_listed_or_removed(monkeypatch):
+    _patch(monkeypatch, "frame.eval_candidate", conflicts=("frame.adapter",), requires=("frame.eval_unrecommended",))
+    _patch(monkeypatch, "frame.eval_default", default=True)
+    _patch(monkeypatch, "frame.eval_unrecommended", recommended=False)
+    _patch(monkeypatch, "frame.eval_inapplicable", applies=False)
+    _patch(monkeypatch, "frame.eval_listed")
+    _patch(monkeypatch, "frame.eval_removed")
+    _patch(monkeypatch, "frame.eval_alternate")
+    _patch(monkeypatch, "adapter.eval_setting", category="adapter")
+    _patch(monkeypatch, "pcvr.eval_wrong_platform", category="pcvr")
+    entry = _entry(frame=["frame.eval_listed"], frame_remove=["frame.eval_removed"],
+                   alt_overport=["frame.eval_alternate"], adapter={"eval_setting": 9})
+    candidates = {r["patch"]: r for r in ev.fit_candidates(_analysis(), entry)}
+    assert "frame.eval_candidate" in candidates
+    assert not set(candidates) & {"frame.eval_default", "frame.eval_unrecommended", "frame.eval_inapplicable",
+                                  "frame.eval_listed", "frame.eval_removed", "frame.eval_alternate",
+                                  "adapter.eval_setting", "pcvr.eval_wrong_platform"}
+    c = candidates["frame.eval_candidate"]
+    assert c["reason"] == "reason for frame.eval_candidate"
+    assert c["category"] == "frame" and c["experimental"] and c["strict"]
+    assert any("conflicts" in w for w in c["warnings"])
+
+
+def test_fit_uses_a32_and_flat_rules(monkeypatch):
+    _patch(monkeypatch, "frame.eval_64_only", a32=False)
+    _patch(monkeypatch, "frame.eval_32_ok", a32=True)
+    candidates = {r["patch"] for r in ev.fit_candidates(_analysis(abis=["armeabi-v7a"]), _entry())}
+    assert "frame.eval_32_ok" in candidates
+    assert "frame.eval_64_only" not in candidates
+    assert "frame.a32_bridge" not in candidates  # implied for 32-bit entries, as engine._a32_predates_bridge does
+    flat = {r["patch"] for r in ev.fit_candidates(_analysis(extra={"vr_kind": "none"}), _entry())}
+    assert "frame.eval_32_ok" not in flat  # needs_vr defaults True, removed by engine._non_quest
+
+
+def test_fit_metadata_dates_and_converted(monkeypatch):
+    monkeypatch.setattr(ev, "fit_candidates", lambda a, e: [dict(patch="frame.new", category="frame",
+                                                               reason="x", experimental=False, strict=False,
+                                                               warnings=[])])
+    entries = {"com.a": _entry(package="com.a", verified={"date": "2026-01-01"}, updated="2026-02-01"),
+               "com.b": _entry(package="com.b", status="unsupported", engine="Other", xr="VrApi")}
+    report = ev.fit_report({"com.a": {"analysis": _analysis(package="com.a").to_dict(), "converted": True}}, entries,
+                           {"frame.new": "2026-03-01"})
+    assert report["games"][0]["candidates"][0]["newer_patch"] is True
+    assert report["games"][0]["entry_date"] == "2026-01-01"  # verified, not updated
+    assert report["games"][0]["converted"]
+    assert report["metadata_only"][0]["package"] == "com.b"
+    assert report["metadata_only"][0]["xr"] == "VrApi"
+    assert report["totals_by_patch"] == {"frame.new": 1}
+    md = ev.render_fit_report(report)
+    assert "metadata only" in md and "com.b" in md and "frame.new" in md
+
+
+def test_snapshot_copied_before_read_and_retried(tmp_path, monkeypatch):
+    source = tmp_path / "snapshot.json"
+    source.write_text('{"complete": true}')
+    calls = []
+    real_copy = ev.shutil.copyfile
+
+    def copying(src, dest):
+        calls.append((src, dest))
+        real_copy(src, dest)
+        if len(calls) == 1:
+            dest.write_text('{"partial":')
+
+    monkeypatch.setattr(ev.shutil, "copyfile", copying)
+    monkeypatch.setattr(ev.time, "sleep", lambda _: None)
+    assert ev.read_snapshot(source) == {"complete": True}
+    assert len(calls) == 2
+    assert all(src == source and dest != source for src, dest in calls)
+    assert json.loads(source.read_text()) == {"complete": True}
+
+
+def test_intro_dates_cached_and_adapter_searches_bare_key(tmp_path, monkeypatch):
+    calls = []
+    patch = base.get("adapter.scene_emul")
+    monkeypatch.setattr(base, "all_patches", lambda: [patch])
+
+    def log(args, **kwargs):
+        calls.append(args)
+        return type("Result", (), {"stdout": "2026-01-02\n2026-02-03\n"})()
+
+    monkeypatch.setattr(ev.subprocess, "run", log)
+    cache = tmp_path / "patch-intro-dates.json"
+    assert ev.patch_intro_dates(cache) == {"adapter.scene_emul": "2026-01-02"}
+    assert '-S"scene_emul"' in calls[0]
+    assert calls[0][-1] == "src/frameport/patches/settings.py"
+    assert ev.patch_intro_dates(cache) == {"adapter.scene_emul": "2026-01-02"}
+    assert len(calls) == 1
+
+
+def test_legacy_dumps_cli_remains_available(tmp_path, monkeypatch, capsys):
+    a = _analysis()
+    snapshot = {a.package: {"analysis": a.to_dict()}}
+    entry = _entry()
+    calls = []
+    monkeypatch.setattr(ev, "read_dumps", lambda p: calls.append(p) or snapshot)
+    monkeypatch.setattr(catalog, "_load_dir", lambda *args: {a.package: entry})
+    monkeypatch.setattr(ev.sys, "argv", ["eval_heuristics.py", str(tmp_path), "--verbose"])
+    assert ev.main() in (0, 1)  # a difference exit is compatible with the original CLI
+    assert calls == [tmp_path]
+    output = capsys.readouterr().out
+    assert "Split: dev" in output and a.package in output
+
+
+def test_snapshot_fit_cli_writes_markdown_and_json(tmp_path, monkeypatch):
+    a = _analysis()
+    snapshot = tmp_path / "synthetic.json"
+    snapshot.write_text(json.dumps({a.package: {"analysis": a.to_dict(), "converted": True}}))
+    md, js = tmp_path / "fit.md", tmp_path / "fit.json"
+    monkeypatch.setattr(catalog, "_load_dir", lambda *args: {a.package: _entry()})
+    monkeypatch.setattr(ev, "patch_intro_dates", lambda p: {})
+    monkeypatch.setattr(ev.quest_dump, "scan", lambda *a: pytest.fail("must not scan with --snapshot"))
+    monkeypatch.setattr(ev.sys, "argv", ["eval_heuristics.py", "--snapshot", str(snapshot), "--fit-report",
+                                        "--out", str(md), "--json", str(js)])
+    assert ev.main() == 0
+    assert "# New patch fit" in md.read_text()
+    assert json.loads(js.read_text())["games"][0]["converted"]
+
+
+def test_evaluate_split_coverage_and_no_snapshot_analysis_mutation():
+    snapshot = {f"com.game.n{i}": {"analysis": _analysis(package=f"com.game.n{i}").to_dict(), "converted": i == 0}
+                for i in range(10)}
+    entries = {p: _entry(package=p) for p in snapshot}
+    before = json.dumps(snapshot, sort_keys=True)
+    reports = {s: ev.evaluate(snapshot, entries, s) for s in ("all", "dev", "holdout")}
+    assert reports["dev"]["total_games"] == 8
+    assert reports["holdout"]["total_games"] == 2
+    assert reports["all"]["total_games"] == 10
+    assert json.dumps(snapshot, sort_keys=True) == before
+    assert reports["all"]["converted_games"] == 1
+
+
+def test_catalog_truth_applies_the_a32_bridge_transform():
+    """A 32-bit game's entry relies on the bridge switching OVRPort and the 64-bit fixes off; the heuristics do the
+    same, so neither side may count those as removals."""
+    a = _analysis(abis=["armeabi-v7a"], engine="Unreal", graphics="GLES")
+    defaults = ev.expected_defaults(a)
+    want = ev.catalog_signature(_entry(), defaults, a)
+    got = ev.signature(engine.suggest(a, use_catalog=False), defaults)
+    assert want["overport_removed"] == got["overport_removed"]
+    assert want["frame_removed"] == got["frame_removed"]
+
+
+def test_dated_score_sets_aside_patches_newer_than_the_entry():
+    want = ev.catalog_signature(_entry(frame=["frame.unity_text_input"]))
+    got = ev.catalog_signature(_entry(frame=["frame.unity_text_input", "frame.hw_video_decode"]))
+    strict = ev.score_signatures(want, got)["headline"]
+    dated = ev.score_signatures(want, got, {"frame.hw_video_decode"})["headline"]
+    assert strict["extra"] and not strict["exact"]
+    assert dated["exact"] and dated["gap"] == {"frame.hw_video_decode": True}
+    assert "frame.hw_video_decode" not in dated["per_patch"]
+
+
+def test_inherited_picks_are_adapter_and_untoggled_device_patches():
+    inherited = ev.inherited_decisions()
+    assert "adapter.haptic_fix" in inherited
+    assert "device.text_input_window" not in inherited  # catalog.TOGGLED_DEVICE: dropped unless the entry names it
+    assert not any(pid.startswith("frame.") for pid in inherited)
+
+
+def test_gap_decisions_use_the_verified_date_not_updated():
+    e = _entry(verified={"date": "2026-09-28"}, updated="2026-10-14")
+    gaps = ev.gap_decisions(e, {"frame.unity_text_input": "2026-10-03", "frame.adapter": "2026-09-28",
+                                "patch_remove_unreal_force_quit": "2026-10-01"})
+    assert gaps == {"frame.unity_text_input", "alt:patch_remove_unreal_force_quit"}
+    assert ev.entry_date(e) == "2026-09-28"

tests/test_games.py

@@ -8,6 +8,12 @@ from frameport.sources import quest_dump
 
 pytestmark = pytest.mark.games
 
+# 2026-10-15 full snapshot (80 dumps, 65 supported catalog games), strict score: 5/65 exact, patch F1 73.8%;
+# rounded down. The dated score (--fit-report's date cache) was 39/65, F1 88.6%. See docs/HEURISTICS_EVAL.md;
+# runtime-only choices and unsupported catalog entries are excluded.
+MIN_HEURISTICS_EXACT_RATIO = 0.07
+MIN_HEURISTICS_PATCH_F1 = 0.73
+
 
 @pytest.fixture(scope="module")
 def dumps():
@@ -36,8 +42,7 @@ def test_every_catalog_game_found_and_detected(dumps):
 
 
 def test_heuristics_reproduce_catalog(dumps):
-    """Without the catalog, heuristics must reproduce the verified recipes (Phantom's use-the-no-ForceQuit-build is a
-    runtime finding the triage makes, not a static one)."""
+    """Ratchet catalog-only static accuracy without requiring runtime findings to be detectable from APKs."""
     import importlib.util
     from pathlib import Path
 
@@ -45,16 +50,15 @@ def test_heuristics_reproduce_catalog(dumps):
     spec = importlib.util.spec_from_file_location("ev", script)
     ev = importlib.util.module_from_spec(spec)
     spec.loader.exec_module(ev)
-    exact = total = 0
+    snapshot = {}
     for g in dumps:
         a = analyze(g.apk, data_bytes=g.data_bytes())
-        entry = catalog.lookup(a.package)
-        if not entry:
-            continue
-        from frameport.recommend import engine
-
-        want, got = ev.signature(engine.suggest(a)), ev.signature(engine.suggest(a, use_catalog=False))
-        want["status_unsupported"] = got["status_unsupported"] = False
-        total += 1
-        exact += want == got
-    assert exact >= total - 1, f"only {exact}/{total} catalog recipes reproduced by heuristics"
+        row = {"analysis": a.to_dict(), "converted": a.is_overport_output}
+        if a.package not in snapshot or snapshot[a.package]["converted"] and not row["converted"]:
+            snapshot[a.package] = row
+    entries = catalog._load_dir(script.parents[1] / "catalog/games", "bundled")
+    result = ev.evaluate(snapshot, entries, split="all")["headline"]
+    assert result["games"], "no supported catalog games found in the dumps"
+    assert result["exact_ratio"] >= MIN_HEURISTICS_EXACT_RATIO, (
+        f"only {result['exact']}/{result['games']} static catalog recipes reproduced exactly")
+    assert result["patches"]["f1"] >= MIN_HEURISTICS_PATCH_F1, result["patches"]

tests/test_heuristics_adapter.py

@@ -0,0 +1,51 @@
+"""Adapter recommendations use the detected runtime path, without a catalog."""
+import pytest
+from test_patches import _analysis
+
+from frameport.patches import base
+from frameport.recommend.engine import suggest
+
+
+@pytest.mark.parametrize(("extra", "other_libs", "converted", "expected"), [
+    # Built-in Oculus has the system-app check; modern Unity OpenXR can still ship OVRPlugin.
+    ({"unity_oculus_check": True}, [], False, True),
+    ({"unity_oculus_check": False}, ["libUnityOpenXR.so"], False, False),
+    ({"unity_oculus_check": False}, [], False, False),
+    # Neither the version nor the graphics API is the reason to select this fix.
+    ({"unity_oculus_check": False, "unity_version": "2019.4.20f1"}, [], False, False),
+    ({"unity_oculus_check": True, "unity_version": "2019.4.20f1"}, [], False, True),
+    # The built-in check can coexist with XR Plugin Management: retain both exclusions.
+    ({"unity_oculus_check": True, "oculus_xr_plugin": True}, [], False, False),
+    ({"unity_oculus_check": True}, ["libOculusXRPlugin.so"], False, False),
+    # Old analyses have no check signal. Converted builds may have the check patched out.
+    ({}, [], False, True),
+    ({"unity_oculus_check": False}, [], True, True),
+    ({"unity_oculus_check": False}, ["libOculusXRPlugin.so"], True, False),
+])
+def test_pose_time_fix_uses_builtin_oculus_evidence(extra, other_libs, converted, expected):
+    a = _analysis(libs=["libunity.so", "libOVRPlugin.so", *other_libs],
+                  extra=extra, is_overport_output=converted)
+    patch = base.get("adapter.pose_time_fix")
+    detected = patch.detect(a)
+    assert patch.applies(a)  # still available for manual investigation in any OVRPlugin game
+    assert bool(detected and detected.recommended) is expected
+    recipe = suggest(a, use_catalog=False)
+    assert (patch.id in recipe.patches) is expected
+    if expected:
+        assert recipe.patches[patch.id] == {"value": 1}
+    # The new evidence is specific to pose timing: OVRPlugin's haptic bug affects both paths.
+    assert recipe.patches["adapter.haptic_fix"] == {"value": 1}
+
+
+def test_pose_time_fix_still_requires_ovrplugin():
+    a = _analysis(libs=["libunity.so"], extra={"unity_oculus_check": True}, is_overport_output=False)
+    patch = base.get("adapter.pose_time_fix")
+    assert not patch.applies(a)
+    assert patch.detect(a) is None
+    assert patch.id not in suggest(a, use_catalog=False).patches
+
+
+def test_unreal_pose_evidence_is_independent_of_unity():
+    a = _analysis(engine="Unreal", libs=["libUE4.so", "libOVRPlugin.so"], is_overport_output=False,
+                  extra={"unity_oculus_check": False, "unreal_thumb_touch": True})
+    assert suggest(a, use_catalog=False).patches["adapter.pose_time_fix"] == {"value": 1}

tests/test_heuristics_unity.py

@@ -0,0 +1,136 @@
+"""Unity heuristics from executable paths and tool support, without catalog/package matching."""
+import pytest
+from test_patches import _analysis
+
+from frameport.patches import base
+from frameport.recommend.engine import suggest
+
+IL2CPP_FIXES = ("frame.unity_text_input", "frame.unity_runtime_msaa_off")
+
+
+def _il2cpp(**changes):
+    values = dict(libs=["libunity.so", "libil2cpp.so", "libOVRPlugin.so"], graphics="GLES",
+                  extra={"unity_version": "2019.4.10f1", "text_fields": ["TMP_InputField"],
+                         "ovr_runtime_msaa": True})
+    values.update(changes)
+    return _analysis(**values)
+
+
+@pytest.mark.parametrize("pid", IL2CPP_FIXES)
+@pytest.mark.parametrize("version", ["2017.4.23f1", "2017.1.0f1", "5.6.7f1"])
+def test_unreadable_old_metadata_does_not_recommend_il2cpp_rewrites(pid, version):
+    a = _il2cpp(extra={"unity_version": version, "text_fields": ["TMP_InputField"], "ovr_runtime_msaa": True})
+    p = base.get(pid)
+    assert p.detect(a) is None
+    assert pid not in suggest(a, use_catalog=False).patches
+    assert p.applies(a)  # still available for an explicit recipe; only automatic detection changes
+
+
+@pytest.mark.parametrize("pid", IL2CPP_FIXES)
+@pytest.mark.parametrize("version", ["2018.4.0f1", "2019.4.1f1", "2022.3.1f1", "6000.5.6f1", None, "unknown"])
+def test_supported_or_unknown_metadata_keeps_il2cpp_recommendations(pid, version):
+    a = _il2cpp(extra={"unity_version": version, "text_fields": ["InputField"], "ovr_runtime_msaa": True})
+    assert base.get(pid).detect(a).recommended
+    assert pid in suggest(a, use_catalog=False).patches
+
+
+@pytest.mark.parametrize("pid", IL2CPP_FIXES)
+@pytest.mark.parametrize("abis", [["armeabi-v7a"], ["x86_64"], []])
+def test_aarch64_return_instructions_are_not_recommended_for_other_abis(pid, abis):
+    assert base.get(pid).detect(_il2cpp(abis=abis)) is None
+
+
+@pytest.mark.parametrize("pid", IL2CPP_FIXES)
+def test_il2cpp_rewrites_still_need_unity_and_il2cpp(pid):
+    assert base.get(pid).detect(_il2cpp(engine="Unreal")) is None
+    assert base.get(pid).detect(_il2cpp(libs=["libunity.so", "libmonobdwgc-2.0.so"])) is None
+
+
+def test_text_classes_remain_a_signal_for_both_unity_field_implementations():
+    p = base.get("frame.unity_text_input")
+    for fields in (["TMP_InputField"], ["InputField"], ["TMP_InputField", "InputField"]):
+        a = _il2cpp(extra={"text_fields": fields})
+        assert p.detect(a).recommended
+        assert base.get("device.text_input_window").detect(a).recommended
+    assert p.detect(_il2cpp(extra={"text_fields": []})) is None
+
+
+@pytest.mark.parametrize("graphics", ["Vulkan (declared in manifest)", "Vulkan (Unity boot.config)"])
+def test_runtime_msaa_does_not_guess_gles_from_a_vulkan_label(graphics):
+    assert base.get("frame.unity_runtime_msaa_off").detect(_il2cpp(graphics=graphics)) is None
+
+
+def test_runtime_msaa_needs_the_runtime_msaa_signal():
+    assert base.get("frame.unity_runtime_msaa_off").detect(_il2cpp(extra={}, unity_msaa_levels=4)) is None
+
+
+@pytest.mark.parametrize("version", ["2017.4.23f1", "2019.4.20f1", None])
+def test_built_in_oculus_path_still_needs_system_app_bypass(version):
+    a = _il2cpp(extra={"unity_version": version, "unity_oculus_check": True})
+    p = base.get("frame.unity_oculus_check")
+    assert p.detect(a).recommended
+    assert p.id in suggest(a, use_catalog=False).patches
+
+
+def test_oculus_xr_plugin_still_gets_the_builtin_startup_check():
+    """A verified game (com.coatsink.alone) ships libOculusXRPlugin.so and still needs the bypass."""
+    a = _il2cpp(libs=["libunity.so", "libil2cpp.so", "libOVRPlugin.so", "libOculusXRPlugin.so"],
+                extra={"unity_version": "2019.4.35f1", "unity_oculus_check": True})
+    assert base.get("frame.unity_oculus_check").detect(a).recommended
+
+
+def test_oculus_metadata_field_alone_does_not_prove_a_native_plugin_path():
+    a = _il2cpp(extra={"unity_oculus_check": True, "oculus_xr_plugin": True})
+    assert base.get("frame.unity_oculus_check").detect(a).recommended
+
+
+def test_oculus_bypass_needs_both_the_check_and_ovrplugin():
+    p = base.get("frame.unity_oculus_check")
+    assert p.detect(_il2cpp(extra={})) is None
+    assert p.detect(_il2cpp(libs=["libunity.so"], extra={"unity_oculus_check": True})) is None
+    assert p.detect(_il2cpp(engine="Other", extra={"unity_oculus_check": True})) is None
+
+
+FIRESTORE = "libFirebaseCppFirestore.so"
+GECKO = "libxul.so"
+
+
+def _tls(version="6000.5.6f1", heavy=(FIRESTORE, GECKO), **changes):
+    return _analysis(libs=["libunity.so", *heavy], extra={"unity_version": version}, **changes)
+
+
+@pytest.mark.parametrize("version", ["6000.0.1f1", "6000.5.6f1"])
+def test_unity6_firestore_and_gecko_recommend_thread_key_pool(version):
+    a = _tls(version)
+    p = base.get("frame.unity_tls_key_pool")
+    assert p.detect(a).recommended
+    assert p.id in suggest(a, use_catalog=False).patches
+
+
+@pytest.mark.parametrize("heavy", [(FIRESTORE,), (GECKO,), ("libFirebaseCppApp-9_6_0.so", GECKO)])
+def test_less_thread_key_pressure_remains_an_optional_diagnostic(heavy):
+    a = _tls(heavy=heavy)
+    p = base.get("frame.unity_tls_key_pool")
+    assert not p.detect(a).recommended
+    assert p.id not in suggest(a, use_catalog=False).patches
+
+
+@pytest.mark.parametrize("version", ["2019.4.1f1", "2022.3.1f1", None, "unknown"])
+def test_thread_key_pressure_rule_requires_unity6(version):
+    assert base.get("frame.unity_tls_key_pool").detect(_tls(version)) is None
+
+
+def test_thread_key_pool_requires_native_unity_for_arm64():
+    p = base.get("frame.unity_tls_key_pool")
+    assert p.detect(_tls(engine="Unreal")) is None
+    assert p.detect(_tls(abis=["armeabi-v7a"])) is None
+    assert p.detect(_analysis(libs=[FIRESTORE, GECKO], extra={"unity_version": "6000.5.6f1"})) is None
+
+
+def test_rendering_tradeoffs_stay_opt_in_without_the_runtime_symptom():
+    a = _il2cpp(libs=["libunity.so", "libil2cpp.so", "libOculusXRPlugin.so"], unity_msaa_levels=4,
+                extra={"oculus_xr_plugin": True, "ovr_runtime_msaa": True})
+    for pid in ("frame.unity_multipass", "frame.unity_no_overlay_copy"):
+        p = base.get(pid)
+        assert p.applies(a) and p.detect(a) is None
+        assert pid not in suggest(a, use_catalog=False).patches