fix(tools): MemoryReviewer --dry-run builds the prompts and stops

[pai-scaffolde]

Reproduced on a fresh LifeOS 7.40.4 install (macOS, Claude Code in the Claude desktop app); the fix was applied to that install and run there before filing. Fixes #2073 (the shape suggested there is the one taken).

Observed

bun LIFEOS/TOOLS/MemoryReviewer.ts review --dry-run called the model, created MEMORY/OBSERVABILITY/reviewer-runs/<runId>/ and appended a row to reviewer-runs.jsonl, while the file's own header (line 28) promises "extract + prompt, no inference".

Root cause

review() threads dryRun into exactly one place, dispatchItems(). Everything upstream of that — writeRunDebug(), inference() (twice on a parse retry), logRunSummary() — runs unconditionally. Those two artifacts are what CortexHealth.reviewerEvidence() grades: a run directory newer than the latest ledger row is treated as an orphaned run and, past the grace window, reported timed-out; every dry run also inflates the reviewer denominators MemoryFixesCheck/MemoryInsights/MemoryStatus read.

Fix (behavior, so the header becomes true)

review() returns right after the prompts are built when dryRun is set: no inference, no run directory, no ledger row. The result carries dry_run: true and prompt_chars (system/user sizes) alongside transcript and exchanges, so the operator sees what would have been sent. dispatchItems() keeps its own dryRun option (the smoke test uses it); review() no longer passes it since a dry run never reaches dispatch. The header line is unchanged because it now describes the code.

How tested

bun MemoryReviewer.ts test (the tool's existing self-test), run under a temp HOME with the USER/MEMORY symlink layout, gains five checks: dry run returns ok with dry_run, inference_duration_ms === 0 and no dispatch_summary, run-dir count unchanged, ledger row count unchanged, prompt sizes reported.

upstream file + the new checks (negative control):  dry-run: no run directory written — 1 → 2 ✗ ; no ledger row written — 1 → 2 ✗
patched:                                              all five dry-run checks ✓

The remaining self-test checks are unchanged before and after. Transpiles. On the live install: review --dry-run leaves the run-dir and ledger counts unchanged, and Cortex health reports ok.

🤖 Generated with Claude Code