perf(kv-unified): pack labels and selectively rescore terminal cuts

[antra-tess]

Consolidated into #110 at 3356af7. GitHub marked this PR merged because its commits are now in #110's branch. Nothing has been merged into main; review and merge #110 only.

Problem

After #110, general solves still spend substantial time allocating/grouping labels and rescanning every leaf of every retained terminal cut. This implements the two follow-up leads: reusable packed propagation storage and selective exact terminal rescoring.

Changes

• Use owned numeric label handles, contiguous binary64 records, immutable paged trace indices, and reusable generation-stamped bucket tables. Sparse fallback changes only storage, never pruning or stopping rules.
• Cache action-level fidelity/continuity sums with conservative floating-point intervals; compute chronological cache costs exactly for every retained cut.
• Establish exact normalization floors independently of welfare, then exactly rescore every possible winner, tie, and carried-layout contender. Preserve the existing comparator and hysteresis rule.
• Preserve the complete sorted candidate API with deferred exact evaluation when the diagnostic list is requested. Keep storage: 'objects' and terminalEvaluation: 'full' as differential-reference modes.
• Add storage/snapshot/bounds/tie/floor tests, per-estimate benchmark verification, replay reference checks, design notes, and a changelog fragment.

Policy weights, grids, hard walls, label ceilings, and adoption thresholds are unchanged. No solve deadline is introduced. All performance measurements disable the certificate.

Tests

• npm run build: pass.
• npm run typecheck: pass.
• Full compiled Node suite: 822 pass / 0 fail, versus 813 / 0 on the fix(kv-unified): accelerate general solves and reduce memory use #110 parent; 9 added tests.
• 90 varied forests: object/full, packed/full, and packed/selective have identical selected candidates, all retained candidates, floors, and pre-existing propagation/envelope statistics. Every feasible small cut is also checked against its bounds.
• Saved 73,918-leaf / 3,345-summary fixture, fresh-process Bun runs:

Case	Object/full	Packed/selective	Reduction
Warm cache, forced 500k wall	8.721 s	5.194 s	40.4%
Warm cache, newly pinned middle leaf	8.736 s	5.229 s	40.1%
Warm cache, 30k appended tail tokens	9.149 s	5.509 s	39.8%

All selected frontier hashes, scores, floors, and existing propagation/envelope statistics match exactly. Only 3 exact evaluations are needed in each case, versus 6,538–6,719 retained cuts (the count includes the feasibility-witness check). A separate diagnostic checked every pinned-case estimate against exact evaluation.

Controlled 500-message replay at 028b065: 500 successes / 0 failures, certificate disabled.

Statistic	Previous replay	Packed/selective	Reduction
Mean	3.122 s	2.044 s	34.5%
Median	2.837 s	1.878 s	33.8%
p95	4.959 s	3.145 s	36.6%
p99	8.069 s	4.927 s	38.9%
Maximum	9.486 s	5.611 s	40.9%

420/500 solves are below 2 seconds, 495/500 below 5 seconds. All 23 exact-reference checks pass (warm-up, every 25th message, and every layout movement), including all four transitions. Checks compare frontier hashes, all selected loss terms, score, and both floors; reference time is outside the solve timer. Every row's tokens, movements, terminal labels, and maximum labels per state also match the old replay. The old rows lack hashes/scores, so this is not a claim of full decision comparison on all 500 historical rows.

The previous replay is the archived pre-rebase 5c55a26 run; a fresh complete #110-parent replay was not repeated. The paired large-fixture comparisons above use the current object backend. Replay timing includes forest/solver construction, solve, and selected frontier materialization; one warm-up is excluded. It models accepted layouts and fresh atomic cache markers using a frozen summary catalogue, not historical provider calls or cache TTLs. Observed high-water RSS was 3.285 GB including the in-process reference solves.

CI passes on macOS and Ubuntu with Node 20 and 24. Full design, reproduction steps, and validation details are in docs/kv-unified-packed-selective.md.

Not verified / out of scope

• No live resident changes, deployment, provider calls, cache-TTL simulation, or historical compression reconstruction.
• Local timings are measurements, not a universal runtime bound or a cross-platform guarantee.
• Diagnostic candidate-list inspection intentionally performs remaining exact scoring; its cost is not eliminated.

Companion PRs

All changes are now incorporated into #110 without rewriting published history. This PR no longer needs separate review or merging.

🤖 Generated with Codex (GPT-6).

[antra-tess]

Consolidated into #110 at 3356af7, preserving both commits and their validation history. GitHub marked this PR merged into its base branch; nothing was merged into main. Please review and merge #110 only.