feat: ship the typed solver with exact optimization and a shared-build test matrix

[Marcus-Rise]

The restored library branch is ready for review against main. This PR makes the solver minimize grouped actions first, then unit shifts among equal-action solutions, and separates verification from release delivery.

• Share one strict TypeScript core between src/index.ts and src/cli.ts. The public contract remains solveLock(state, links, config?) with a documented createSolverConfig(overrides?) factory.
• Build four self-contained library/browser files plus one standalone Node CLI, with declarations and no runtime dependencies.
• Keep source flat, tests and benchmarks under tests/, and maintained documentation under wiki/. Generated reports, snapshots, coverage, traces and builds are CI artifacts.
• Remove .github/scripts and its wrapper-only tests. pnpm build invokes Vite and TypeScript directly; CI and publication commands stay in two flat YAML workflows.
• CI builds and benchmarks once, then tests the same archive in six matrix rows: Node 22/24/26 and Chromium/Firefox/WebKit. The browser rows use Node 24 to drive tests. Release publishes a canary only after the entire main-push CI succeeds; stable publication remains a manual dispatch selecting a successful main CI run.
• Keep build/release jobs short: use the runner's official GitHub CLI, native run/artifact queries and standard publication commands. Remove manual CLI installation, custom run traversal, duplicate runtime assets, repeated release-file validation, checksum assets and the CDN polling gate.
• Canonical artifacts are named by SHA within their CI run. Native failed-job reruns reuse the successful build; full build reruns overwrite its artifacts through the official Action.
• Compare benchmarks only with a saved report for the exact target SHA. The latest successful CI run for that SHA supplies the report. Missing or expired evidence is explicitly skipped; no target checkout, rebuild, registry or release fallback.
• Repack only version metadata for release. Tested runtime bytes are preserved. Publication uses direct npm pack, npm publish and gh release create commands. Release notes include npm/CDN links and the benchmark report.

The A* queue and relaxation now compare both costs. The matrix certificate and dense/sparse BFS fallback honor the same ordering. An equal-action route with fewer shifts can improve the predecessor before the result is finalized. Domain validation, intermediate-position bounds and search budgets remain enforced. The proof, complexity and limits are documented in the algorithm Wiki.

The reported regression is fixed:

Lock	Actions before / after	Unit shifts before	Unit shifts after
lock-011	9 / 9	41	37
lock-018	14 / 14	49	45
lock-019	14 / 14	43	41
All 45	483 / 483	1723	1713

These before/after values compare the saved pre-fix solver with the final built module. No lock increased its actions, shifts, distinct plates or switches. All 45 final (actions, shifts) pairs were also checked against an independently implemented exhaustive solver. Focused regression tests and an independent exhaustive oracle for small state spaces are maintained in the test suite. Minimum distinct-plate count is not an API guarantee.

CI passed on the exact head, c6662611494f2b2181329b16dc212f5e9b84be7f:

• 237 Node tests in each of the Node 22/24/26 rows; 42 tests of the deleted infrastructure wrappers were removed. Source coverage is unchanged. The installed archive and CLI pass in every row.
• 543 tests in real Chromium, Firefox and WebKit; 13 distribution tests; 24 static-browser/Worker consumers and 24 installed-package browser consumers.
• Coverage: 96.20% statements, 92.38% branches, 100% functions, 96.01% lines; the minimum is 80% for each metric.
• Lint, strict type checking, all five bundles and the 45-lock benchmark passed.
• Canary and stable preparation were additionally exercised locally with npm 12.0.2; only version metadata changes, and packaged runtime bytes stay identical.

The current main SHA a87e739a22ccc4215d6d9b14fa06b0738171cb60 has no saved report matching the new artifact contract, so this CI run skipped historical comparison. That skip is not evidence of equivalence; the regression result above comes from the separate pre-fix and independent-oracle checks. Future runs compare with available exact-SHA reports. Historical timing/RSS ratios remain informational because machines and runtime conditions may differ.

The benchmark artifact contains the complete JSON report, commands, raw samples, environment and Markdown table. The build artifact contains the tested package (including all five runtime files) and the same JSON/Markdown evidence. Downloaded artifact digests, report source identity and packaged file hashes were verified. All six matrix rows verified the same npm archive SHA256: b1517f701ee7cdf221fe284014b1c9b501007721715e488a1795ab494788a8dd.

All 45 locks from the final CI report

Node 24.19.0, Linux x64, AMD EPYC 7763; two warmups and seven measured solves per lock. Timings cover the solver call only, excluding cloning and independent validation.

Lock	Actions	Distinct plates	Unit shifts	Switches	Median ms	p95 ms
lock-001	17	6	57	16	5.767	9.435
lock-002	9	5	31	8	0.129	0.214
lock-003	5	4	15	4	0.060	0.082
lock-004	14	5	60	13	0.383	0.395
lock-005	11	6	41	10	0.359	0.754
lock-006	9	5	30	8	0.140	0.153
lock-007	9	6	28	8	0.237	0.260
lock-008	5	4	16	4	0.033	0.035
lock-009	17	6	72	16	3.163	3.796
lock-010	7	5	25	6	31.396	41.904
lock-011	9	6	37	8	0.143	0.229
lock-012	23	5	68	22	2.195	2.436
lock-013	20	6	75	19	9.622	12.181
lock-014	9	7	38	8	0.118	0.145
lock-015	7	5	31	6	0.118	2.546
lock-016	11	5	44	10	0.523	0.781
lock-017	13	6	52	12	0.742	1.292
lock-018	14	6	45	13	9.271	13.218
lock-019	14	5	41	13	2.212	2.269
lock-020	12	6	56	11	0.056	0.092
lock-021	11	6	35	10	2.177	5.691
lock-022	8	6	23	7	0.123	0.165
lock-023	16	6	48	15	7.207	15.880
lock-024	8	5	29	7	0.320	0.365
lock-025	14	5	44	13	1.021	1.034
lock-026	9	5	37	8	0.082	0.122
lock-027	8	5	36	7	0.082	0.110
lock-028	11	6	43	10	0.726	0.751
lock-029	20	6	48	19	11.863	14.151
lock-030	10	6	41	9	0.186	0.211
lock-031	9	5	30	8	0.148	0.193
lock-032	9	6	31	8	0.102	0.151
lock-033	10	5	31	9	0.273	0.298
lock-034	8	6	33	7	0.087	0.099
lock-035	10	6	40	9	0.098	0.179
lock-036	8	5	27	7	0.075	0.106
lock-037	11	6	40	10	0.218	1.836
lock-038	8	6	21	7	0.306	0.338
lock-039	8	6	25	7	0.098	0.117
lock-040	13	5	46	12	1.093	1.190
lock-041	9	5	36	8	0.032	0.048
lock-042	4	3	20	3	0.014	0.044
lock-043	5	5	13	4	0.015	0.017
lock-044	11	6	41	10	0.309	0.336
lock-045	10	5	33	9	0.538	0.670

Totals: 483 actions, 1713 unit shifts, 438 switches, and 245 distinct plates summed per lock. Whole-catalog process peak RSS: median 120,549,376 bytes across three fresh processes. This includes Node, loading and validation; it is not solver-only allocation or browser memory.

Reproduce with Node 24.19.0:

npm install --global --ignore-scripts npm@12.0.2 pnpm@12.3.4
pnpm install --frozen-lockfile
pnpm exec playwright install --with-deps chromium firefox webkit
pnpm test
pnpm benchmark

To compare a downloaded report for a known exact target SHA:

pnpm benchmark --baseline-report /path/to/benchmark.json --baseline-sha <target-sha>

Release setup and the manual stable command are documented in the release Wiki. The PR remains a draft targeting main; no merge or npm publication has been performed.