perf: prefilter behavior shortcuts for larger catalogs

[yeahjack]

Summary

Prefilter behavior shortcuts with the existing pure press/release matchers when the active catalog has at least 32 behaviors. Unrelated inputs avoid scanning behavior IDs. Smaller catalogs retain ID-first matching because parsing stale bindings can cost more than a short scan.

The production optimization is now only 8 added lines. The original inner ID-first matching/action body and following code remain verbatim; actual large-catalog hits intentionally recheck the pure matchers. Binding/catalog order, global priority, repeated downs, releases, expression selection and Alt+1…0 fallback are unchanged. No cache, allocation or persistent state is added.

Independent base: 4283de1599c7da914138f82a99405d67c2c861ec. This refresh contains only this optimization, its validation and the small build cleanups below; it does not restore functionality removed upstream or depend on another optimization PR.

Verification

• Full Linux Release diagnostic build with warnings as errors; all 6 CTests pass
• 16,590 ordered external-action and complete-app-state comparisons against a frozen copy of current upstream, verified byte-for-byte apart from test symbol remapping
• Explicit expected action traces and seeded differential sequences cover the 31/32/33 boundary, duplicate/stale/empty IDs, failures, priorities, releases, repeated downs, all Alt digits, gamepad thresholds and malformed shortcuts
• Focused ASan/UBSan passed; source line policy and diff whitespace checks passed
• The acceptance contract was frozen before gap-closing changes: Acceptance spec and traceability map SC-01…09 to code, tests and limits
• Spec-driven tests cover an uninterrupted held-input stream through 31↔32↔33, zero/regrowth, in-place binding/catalog edits and fallback; every timed fixture now passes ordered-trace/full-state parity preflight
• Independent review checked final production semantics, test-oracle provenance, assembly, build flags and every published statistic; injected POSIX clock failure exits clearly without measurement rows

Fresh measurements on this base

GCC 14.2, Release -Os plus IPO/LTO, Xeon Platinum 8573C shared Linux cloud host. Old/new implementations use the same executable, compiler settings and validator-checked fixtures. Three runs each contain three warmup pairs and 31 alternating measured pairs per scenario. Other project builds/tests were paused; no exclusive CPU or affinity isolation is claimed.

Medians of the three run-level wall medians:

Scenario	Before	After	Median speedup range
128 behaviors, 8 mixed bindings	5.820 µs/event	1.476 µs/event	3.87–7.04×
128 behaviors, 128 unmatched bindings	87.684 µs/event	9.364 µs/event	9.30–10.52×
128 label overrides, 8 active shortcuts	89.555 µs/event	1.793 µs/event	41.32–85.86×

The first cached-match version caused an avoidable stale-31 regression. GCC/LTO inspection showed loop indices spilling to memory while match booleans occupied registers. The simpler guard restores both loop indices to registers. In three contemporaneous cached/guard comparisons, stale-31 changed from 0.884–0.943× to 1.094–1.687×, without tuning the threshold or fixtures. This supports removing the structural cost; noisy timings do not establish its precise causal magnitude.

Residual costs remain: one behavior with 64 stale bindings costs 6.60–37.01 ns/event more (0.9–7.2%) in these runs. The guard adds a decision per binding and the measured dispatcher stack reservation is 72 bytes versus baseline 56 (rejected cached version: 88); these are compiler-specific benchmark observations, not application memory claims. The threshold is an empirical tradeoff, not a universal speedup guarantee. Run 2 has substantial shared-host variation, so upper speedup ranges should not be treated as stable device performance.

All 78 scenario/run summaries and 2,418 final guard-only raw paired samples, wall/process-CPU measurements and reproduction commands are in tests/runtime/README-shortcuts.md and its linked CSV files. These are dispatch component costs with deterministic external-action doubles. Batch-average p95 values do not establish individual-event tails, whole-app CPU, FPS or input-to-frame latency. Bounded calibration and shared-host noise limit small-difference conclusions. No previous-base measurements are reused.

Build compatibility and scope

Two small fixes for warnings inherited from the current baseline are kept separate from the optimization: split two independent audit if statements onto separate lines, and compile remove_receipt under the same BONGO_CAT_HAS_CUBISM condition as its existing caller. The helper body and call path are unchanged. No warning exemptions are required.

Linux diagnostic-backend coverage does not establish native Windows/macOS behavior, licensed Cubism rendering, GPU/window correctness or input-to-photon latency. LeakSanitizer is unavailable under this executor's tracing; leak checking is not claimed.

Current-head review and CI status

Published head: 758fa1294b2b169c8b8455ada637d45e86c802e2; tested tree: 63c6bb13e6de74ea1236d12ffd974c6d10202c69. The tree was verified against the tested local source after a protected, expected-old-head ref update. GitHub reports no merge conflict and the PR is Ready for review. No merge, auto-merge or deployment was performed.

CI has not passed: the current-head workflow run completed as action_required, with zero executed jobs and zero check runs. A repository maintainer must approve the workflow before cross-platform jobs can run. The fork has no run for this head. Local strict builds and focused tests are the established evidence; Ready for review is not a ready-to-merge claim.