perf(ci): add focused benchmark commands

[marcus-pousette]

Why

The exhaustive benchmark workflow takes about 57 minutes. That is useful before a performance-sensitive merge, but too expensive for routine PR feedback.

Most of the time is real benchmark coverage rather than a second benchmark framework: the workflow orchestrates the existing web, native SQLite, direct-sync, Wasm, and Rust-core runners. This PR adds a focused mode that samples those same implementations at representative sizes.

Commands

• /bench or /bench quick — focused cross-runtime signal (27 result rows)
• /bench web — the existing full browser suite, including browser sync
• /bench full — the existing exhaustive cross-runtime suite

Manual workflow runs expose the same quick, web, and full choices, with quick as the default.

What quick keeps

• Browser memory and OPFS operation coverage
• Native SQLite operation and note-path coverage
• Native direct-sync coverage
• Wasm and Rust-core coverage

The browser-specific sync matrix remains in /bench web. This avoids making old base/head revisions use different browser instrumentation while keeping a fast sync signal through the existing native direct-sync runner.

Other cleanup

• Checks out the small quick orchestrator from the trusted workflow revision, so it can benchmark older open PRs that do not contain the new script yet.
• Avoids a duplicate successful aggregation pass.
• Removes unnecessary full-history checkouts and unused Rust components.
• Reuses the existing native-SQLite package-script alias in the full command.
• Documents how the package-local and root-level runners relate.

The focused suite is not a replacement for /bench full; it is the routine feedback path.

Measured impact

The exhaustive run took 56m47s. The validated focused run took 16m42s, a roughly 71% wall-time reduction.

Focused-run breakdown:

• Full setup and two project builds: about 8m50s
• Base measurements: 3m48s
• Head measurements: 3m48s
• Results: 27 rows per revision, with identical groups

Build/setup is now the largest remaining fixed cost. The companion iteration fix in #215 removes additional hidden native oversampling.

Validation

• Prettier, Node syntax, Rust formatting, and git diff --check
• Verified pnpm forwards the browser filter through both package-script levels
• Playwright test discovery confirms quick selects exactly the two generic browser tests
• Independent diff review: no blockers; expected result count is 27
• Live workflow run: success in 16m42s
• Artifact contains 27 base rows and 27 head rows; PR comparison comment updated successfully

Merged correctness/performance fix: #215. Its benchmark regression tests remain in the native test command after the rebase.

[github-actions]

Benchmarks

Triggered by workflow_dispatch - Run: https://github.com/cybersemics/treecrdt/actions/runs/29268135321
Suite: quick - Base command: success - Head command: success
Base: main (683b623) - Head: perf/benchmark-focused-commands (1b34e70)
Compared entries: 27 - Improved: 18 - Regressed: 9

Impl	Storage	Workload	Base ops/s	Head ops/s	Δ ops/s	Base p50 ms	Head p50 ms	Δ p50	Status
core-inmem-crdt	memory	insert-move-1000	13829.97	13789.78	-0.29%	144.61	145.03	+0.29%	✅
sqlite-node	file	insert-into-balanced-tree-fanout10-10000	478.27	473.72	-0.95%	2.09	2.11	+0.96%	✅
sqlite-node	file	insert-move-1000	18292.93	19410.90	+6.11%	109.33	103.03	-5.76%	🟡
sqlite-node	file	read-children-payloads-fanout10-10000	14322.21	14947.32	+4.36%	0.77	0.74	-4.18%	✅
sqlite-node	file	sync-balanced-children-cold-start-fanout10-1000	355.03	385.26	+8.51%	28.17	25.96	-7.84%	🟡
sqlite-node	file	sync-balanced-children-payloads-cold-start-fanout10-1000	635.62	691.56	+8.80%	33.04	30.37	-8.09%	🟡
sqlite-node	file	sync-one-missing-1000	10.31	10.73	+4.02%	96.95	93.21	-3.86%	✅
sqlite-node	memory	insert-into-balanced-tree-fanout10-10000	1416.28	1432.13	+1.12%	0.71	0.70	-1.11%	✅
sqlite-node	memory	insert-move-1000	17434.72	17674.00	+1.37%	114.71	113.16	-1.35%	✅
sqlite-node	memory	read-children-payloads-fanout10-10000	22464.20	22312.24	-0.68%	0.49	0.49	+0.68%	✅
sqlite-node	memory	sync-balanced-children-cold-start-fanout10-1000	344.94	398.17	+15.43%	28.99	25.12	-13.37%	🟠
sqlite-node	memory	sync-balanced-children-payloads-cold-start-fanout10-1000	639.51	698.70	+9.26%	32.84	30.06	-8.47%	🟡
sqlite-node	memory	sync-one-missing-1000	10.06	10.44	+3.76%	99.37	95.77	-3.63%	✅
wa-sqlite	browser-memory	insert-chain-100	3003.91	2846.98	-5.22%	33.29	35.13	+5.51%	🟡
wa-sqlite	browser-memory	insert-chain-1000	400.87	383.30	-4.38%	2494.57	2608.92	+4.58%	✅
wa-sqlite	browser-memory	insert-move-100	5148.67	5462.99	+6.10%	38.84	36.61	-5.75%	🟡
wa-sqlite	browser-memory	insert-move-1000	7784.98	7433.70	-4.51%	256.91	269.04	+4.73%	✅
wa-sqlite	browser-memory	replay-log-100	3081.90	3004.36	-2.52%	32.45	33.29	+2.58%	✅
wa-sqlite	browser-memory	replay-log-1000	402.21	382.07	-5.01%	2486.27	2617.31	+5.27%	🟡
wa-sqlite	browser-opfs-coop-sync	insert-chain-100	1156.17	1299.17	+12.37%	86.49	76.97	-11.01%	🟠
wa-sqlite	browser-opfs-coop-sync	insert-chain-1000	375.02	380.37	+1.43%	2666.52	2629.04	-1.41%	✅
wa-sqlite	browser-opfs-coop-sync	insert-move-100	1928.55	2073.83	+7.53%	103.71	96.44	-7.01%	🟡
wa-sqlite	browser-opfs-coop-sync	insert-move-1000	3667.24	4055.52	+10.59%	545.37	493.16	-9.57%	🟠
wa-sqlite	browser-opfs-coop-sync	replay-log-100	1276.81	1307.15	+2.38%	78.32	76.50	-2.32%	✅
wa-sqlite	browser-opfs-coop-sync	replay-log-1000	374.74	381.40	+1.78%	2668.50	2621.92	-1.75%	✅
wasm	memory	insert-chain-1000	26345.45	28481.85	+8.11%	37.96	35.11	-7.50%	🟡
wasm	memory	insert-move-1000	27417.27	27045.18	-1.36%	72.95	73.95	+1.38%	✅

Status: ✅ <5% change | 🟡 5–10% | 🟠 10–30% | 🔴 >30%

Artifacts:

• base summary: bench-base/benchmarks/summary.json
• head summary: bench-head/benchmarks/summary.json
• command logs: bench-base/benchmark.log, bench-head/benchmark.log

[marcus-pousette]

/bench