fix(input): input, inspect and tracking correctness (review batch 3)

[Maxaubert]

Closes #421. Batch 3 of the 2026-10-09 review: input, Inspect and tracking correctness.

Fixed

• 2: gain learner skips ticks where the pointer is pinned at the clip bound (Wind-placed moves are already excluded via lastSetVirtual).
• 3: monitor retarget: when the render half refuses but the transform half accepts (multi-GPU), the session is forced onto the transform engine (renderStale) instead of mapping a render overlay left on the old monitor.
• 4: monitor retarget/re-detect now also runs when Inspect is toggled at 1x.
• 5: game-inspect focus-stealer window is hidden in EndGameInspect.
• 46: stale swallow records cleared by every non-swallowed DOWN (mouse buttons, Inspect commit, bound keys).
• 50: Wind's own mask keystroke (VK 0xE8) no longer stamps the any-key clock (hook and Raw Input).
• 51: g_maskedThisHold cleared on keyboard hook suspend/reinstall.
• 52: click quiet period stamped from Raw Input button edges, not only per-tick GetAsyncKeyState.
• 53: Inspect pans with the measured GainLearner::gainFor; mouse_ballistics, cookPacket, drainCooked, cookActive and their test deleted.
• 55: EditContext (VS Code) with no MSAA caret reports not-found instead of the start-of-line UIA rect.
• 56: line-end ghost suppression is no longer sticky: a suppressed report that persists 300 ms is followed (pure helpers + tests).
• 57: Win32 caret rect normalised with min/max (mirrored/RTL windows).
• 58 (partial): bridge DLL exports are probed with DONT_RESOLVE_DLL_REFERENCES (no DllMain) before the real load.
• 60: a late FocusLookup answer is accepted only for the focus generation it was asked under.
• 64: selection rect uses the last line's rectangle so Shift+arrow growth moves the followed point (pure LastRectOffset + test).
• 66: FocusTracker::run creates its queue before publishing tid_; stop() sets a flag run() checks, so an early stop cannot hang on join.

Skipped

• 54: raw-input UP check/clear is a microsecond window; making it atomic across the two per-key records needs a lock in the LL hook path, a worse trade than the race (self-heals on the next press).
• 58 (rest): signer allowlist and narrowing the WM_COPYDATA filter need field verification against every real JDK vendor and cannot be done safely here; the filter cannot be restricted per process.
• 57 (rest): direction-bound filters were not verified by the review; only the Win32 normalisation was.
• 2 (injected moves): already excluded, curDx is measured against lastSetVirtual.

Tests

• build.bat test: [doctest] test cases: 593 | 593 passed | 0 failed | 0 skipped
• build.bat: builds clean (pre-existing C4996 fopen warnings only)
• build.bat uiaccess: builds clean
• Not run: Playwright (batch 7 only), hands-on field checks of Inspect speed (item 53 changes Inspect pan to the learned gain; field-verify).

No version bump: bumped at merge time.

🤖 Generated with Claude Code

https://claude.ai/code/session_011zPmivSAQeivGUdBjTsaMu

[Maxaubert]

Review follow-up: item 54 is deliberately skipped (a microsecond window; closing it needs a lock in the LL hook path, and a missed clear self-heals on the next press). Item 58 is partial: only the DllMain-free export probe is done; the signer allowlist and narrowing the WM_COPYDATA filter need field verification against real JDK vendors. Both are listed in the PR body. No code change in this pass.

🤖 Generated with Claude Code

https://claude.ai/code/session_011zPmivSAQeivGUdBjTsaMu