[devin-ai-integration]

Summary

Targets two related concerns on the lurch-strafe (Spam) pipeline that you raised: (1) performance of the input → spam-thread hot path, and (2) strict key-state tracking to eliminate ghost inputs and instability.

Performance — lock-free hot path

• Lurch state is now a single atomic. g_activeSpamKeys (std::vector<int>), g_activeKeysMutex, and g_spamKeysEpoch collapse into one std::atomic<uint32_t> g_lurchState (low 4 bits = WASD bitmask, high 24 bits = epoch). Hook thread and spam thread no longer take a mutex to publish/consume the active key set. The publisher uses a CAS loop that only bumps the epoch when the key bits actually change.
• Backend status counters are atomic. KbdHookBackend::statusMutex_ is gone — the LL-hook callback path is now mutex-free, eliminating a real risk of Windows silently unregistering the hook under contention (it enforces a 300 ms timeout by default). Same change applied to InterceptionBackend.
• WASD scancode + flag cache in SpamLogic. SendKeyInputBatch no longer calls MapVirtualKeyW per cycle; the per-VK scan code and KEYEVENTF_EXTENDEDKEY template are cached on first use.
• Stack-allocated INPUT buffer. SendKeyInputBatch uses std::array<INPUT, 4> (WASD bound) instead of a heap vector, removing per-cycle allocator churn.
• KeybindManager state tracker moves from std::map<int,bool> + std::mutex to std::array<std::atomic<bool>, 256>. Same observable semantics, no allocation, no global lock per keystroke.

Strict key-state tracking — ghost-input & instability

• Edge dedup at dispatch. HandleFeatureKeyEvent now short-circuits any event that does not change physicalKeyDown. This drops the Windows autorepeat storm (1 DOWN per ~30 ms while a key is held), which previously drove the full feature chain (axis mutex, publish, CV wake) on every repeat. Behaviour is preserved — every downstream consumer already deduped autorepeats internally.
• Per-VK debounce. New Config::DebounceUs (debounce_us in config.cfg, default 500 µs). Opposite-edge transitions arriving within the window of the previous edge are dropped as bounce / driver echo. The default sits below human dexterity (~10 ms) and above typical mechanical-switch debounce (~50 µs). Set to 0 to disable.
• Memory ordering audit. Acquire/release pair on physicalKeyDown, g_isSpamActive, and g_lurchState so observers on the spam thread / GUI thread can never see a partial state on weakly-ordered architectures (latent on x86, real on ARM).
• Re-validation before SendInput. Spam thread re-checks the lurch epoch and g_isSpamActive immediately before injecting. Closes the snapshot→inject race that previously emitted ghost DOWN events for just-released keys.
• CleanupSpamState uses the same CAS protocol as the publisher; the two operations interleave safely instead of fighting over the vector.

Scope intentionally NOT in this PR

• Focus-loss reconciliation (reconcile physicalKeyDown against GetAsyncKeyState on WM_ACTIVATEAPP / idle tick).
• Synthetic-echo defense (per-VK pending counters to ignore reflected injections from co-resident drivers).
• Logger off the hook thread (queue + worker thread for Logger::Log).
• High-resolution waitable timer replacing timeBeginPeriod(1) + busy-spin in SpamLogic.

These are behaviour-changing or larger refactors; happy to follow up with separate PRs once you confirm direction.

Review & Testing Checklist for Human

• Lurch behaviour in-game. Hold A, hold both A+D, swap which is "winner" (SnapTap on), release while spamming. Spam should not emit ghost DOWN after release; the previously-active key should not stay stuck in the target app.
• Debounce default. Confirm 500 µs default doesn't suppress real fast taps. If you ever notice missed inputs on rapid double-taps, set debounce_us = 0 in config.cfg and retry — that isolates whether debounce is the culprit.
• Backend switch. Toggle between KbdHook and Interception while the spam trigger is active; spam keys should release cleanly across the switch.
• Stats panel in GUI. "Spamming: …" should still reflect the current active set when the trigger is held.
• Rebind flow. Bind the spam trigger to a new key while a key is already held; the new bind shouldn't immediately fire spam.

Notes

• Risk: yellow. The hot path is restructured but every public API and config key is preserved; g_activeSpamKeys was internal-only. The atomic-bitmask layout is documented inline in Globals.h.
• KeyState's deleted copy/move ops are a small correctness improvement: the struct was only ever used as elements of g_KeyInfo[256], so disallowing copy/move is closer to intent and removes the maintenance burden of keeping the manual copy/move ops in sync with new fields like lastEdgeNs.
• The analysis document with the full tier breakdown (Tiers 3 / 4 deferred) is attached to the originating session.

Link to Devin session: https://app.devin.ai/sessions/aefc79c6e4004edfac6d5aef0487ec4b
Requested by: @fL0P1337

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment and CI monitoring