revert: restore the pre-#155 SQLite lock behavior

[brentyates-swx]

Refs #155
Refs #162

This reverts 8cff3506 (#155, "fix(core): restore the SQLite lock-wait window") on swxtch.

Why. Under concurrent writers, 8cff3506 drops more accepted prompts than its parent bf5180afa0 (#150). The report attributes the extra losses to SQLITE_BUSY_SNAPSHOT: #155 excludes that code from retry, so a stale-snapshot failure fails its prompt after the 204 has been returned. The diagnostic build behind that attribution is in the report and was not repeated for this PR. This PR restores the parent's behavior. The fix-forward is tracked in #162, and this PR does not attempt it.

Tree identity

• The revert commit a0703354: git diff --stat bf5180afa0 a0703354 is empty.
• bf5180afa0^{tree} = a0703354^{tree} = 164199bcea117031172486bbe1a0c5bccaedff10.
• The follow-up commit a97b0577 changes only CHANGESET.md, adding this PR's entry.

Multi-writer comparison

Four opencode serve processes per set, one scratch DB per set, heavy load (8 rounds x 24 sessions per server x 32 threads; 768 accepted prompts per run). Short = accepted prompts (HTTP 204) with no persisted user message.

Run	Build	Condition	Short	Lock-failed lines
mw1-R	this revert, a0703354	concurrent with installed	143	222
mw1-I	installed, 86622785	concurrent with revert	190	311
mw2-R	this revert, a0703354	concurrent with parent	151	220
mw2-P	parent, bf5180af	concurrent with revert	147	236
mw3-R	this revert, a0703354	alone	133	230
mw3-P	parent, bf5180af	alone	158	213

• No run reaches the 357 to 389 short band the report measured for 8cff3506. That build was not re-measured here.
• Every lock-failed line carries a SQLITE_BUSY or LockTimeoutError cause.
• Absolute levels run above the report's parent numbers (96 to 111) in this environment. The comparison that holds is reverted against parent in matched runs.
• Fresh-store proof: integrity ok, incremental auto-vacuum, WAL, 41 of 41 migrations.

Tests

• packages/core: sqlite-busy, 20 pass, 0 fail.
• packages/opencode: lock consumers (httpapi-error-middleware, processor-effect), 36 pass, 0 fail.

Changeset

a97b0577 adds this PR's entry under "Fork changes" in CHANGESET.md. script/changeset-check.ts passes for #163 with CI's environment.

Full evidence: data/opencode-revert-155/report.md in the fleet data directory (not in this repo).

🤖 Generated with Claude Code

[brentyates-swx]

Session stamp (claim record)

• Session ID: 383579a1-fdcd-4b9c-9df4-00230aff89bd
• Claude session: https://claude.ai/code/session_019PSxF9tfYzQesA13oxDTHe
• Working repo (launch cwd): /home/byates/fm-worktrees/fm-llmrouter/opencode-revert-155/swx-opencode