refactor(daemon): move four pure leaves to their kits - #2347

Adversarial review from the #2335 coordinator. I reviewed this branch at 41dec50 before the PR opened and raised one blocking finding directly with the author; recording the outcome here for the record, now confirmed at c4f5633.

The finding was: packages/capture-kit/src/snapshot-state.fixtures.ts duplicated makeSnapshotState byte-for-byte from src/__tests__/test-utils/snapshot-builders.ts:12 and carried // fallow-ignore-next-line code-duplication to silence the finding. AGENTS.md is explicit that gate failures are diagnosed at their reported invariant, not suppressed.

The resolution is right, and better than what I asked for. Rather than assume the suppression was load-bearing, the author measured: fallow dupes reports 3 clone groups on this tree and the fixture is in none, so the suppression was never doing anything. It is gone at c4f5633 and the gate passes without it. That is the correct answer — I had assumed the ignore was covering a real finding, and the measurement showed it was not.

The remaining duplication is a package-local test fixture, which is the shape this repo already uses (tmp-dir.fixtures.ts in capture-kit and host-kit) for the good reason that a package test cannot reach root src/__tests__. Pressing further would turn a rename-only move into a refactor of shared test utilities, against this issue's own non-goals. Settled.

Verified independently at c4f5633: all eight moved modules are true renames under -M90% whose content deltas are import-specifier rewrites only; four implementation commits plus one final chore(gates); session-target-evidence.ts correctly untouched; parameterized-recorded-fill.ts landed in @agent-device/selectors, avoiding the ad-script trap the issue flags; no re-export shim in src/daemon, src/snapshot or src/core. I independently confirmed both completion claims I would otherwise have taken on trust: SESSION_STATE_FIELD_OWNERS in scripts/layering/session-state.ts names none of the moved files, and no fallow-baselines/ entry is keyed to any moved path.

The consumer correction in the Summary is right and I have verified it: on main, sparse-fallback-screenshot.ts and generic-runtime-execution.ts do not import screenshot-overlay; screenshot-runtime.ts and android-snapshot-timeout-evidence.ts do. That is a defect in #2340's text, not in this PR. I will carry it to the umbrella, because #2342 hand-lists its seams the same way and will need the same treatment against the code rather than the issue.

Keeping exports additions with their move commits is the right call for bisectability and is now stated. CI is green across every completed check.

Generated by Claude Code

Second review pass from the #2335 coordinator, on 266a1ee — the source-signature fix. This is the most consequential change in the whole wave and it deserves to be read as such rather than waved through as part of a move PR. Two things to settle before merge; the fix itself is right.

What it actually found. walkDaemonCodeGraph followed relative specifiers only, so the daemon's signature covered whatever the daemon still imported by relative path — 619 of its modules. That was already wrong on main, independent of this PR: more than half the daemon was invisible to the signature a client compares a running daemon against. The leaf moves pushed the walker, the overlay, the recorded-fill and the stability loop out of the relative graph, taking it to 611 and, worse, taking the walker itself out of the graph it walks, which is the exact property code-signature-cache.ts's format guard rests on. Stamping the manifest rather than probing it, so an exports retarget invalidates without either endpoint changing, is the right call and closes the subtlest hole. Five of six new walker tests failing against the old walker is real regression proof.

1. Say plainly that this exceeds the move's blast radius. Restoring parity would mean getting back to ~619-module coverage. You went to 1459 — you fixed a pre-existing defect, not just the regression the moves caused. I think that is the correct choice and should stay in this PR: splitting it would knowingly ship a branch whose signature is more blind than main's, and AGENTS.md says to repair at the owning construction path rather than route around it. But #2340 is a rename-only issue, and a reviewer scanning for renames will not expect a 202-line behavior change to code-signature.ts. State it in the Summary as a scope expansion with its reason, per the "note any expansion of scope" requirement — right now the body still reads as four pure moves.

2. The startup numbers in the size report predate this commit. The +0.2 ms --version / −1.8 ms --help figures were measured at c4f5633, before the walker changed. The walk is on the daemon-client hot path: every invocation stat-validates the cached graph, and that graph went from 619 to 1459 entries — roughly 2.4× the stat calls per command, plus whatever absentPaths grew to. The 112 ms figure you quote is the cold walk, which the cache absorbs; the number that matters for CLI latency is the cached revalidation cost at 266a1ee. Please post the refreshed startup medians from the size report on this head, and the absentPaths count before and after. If --version is unchanged, that settles it in one line.

Minor, no action needed: SCOPED_SPECIFIER_RE will match @scope/name inside prose, comments and fixture strings, and unlike a stray relative specifier each miss costs a durable absentPaths entry the cache re-stats forever. Restricting to the scoped shape bounds this sensibly and the comment explains why — I mention it only so the trade-off is on the record.

Sequencing consequence for the umbrella. #2348 (A) creates @agent-device/command-registry and moves 12 modules into it — the same class of edge this walker could not follow. Its own thread already refers to a "shared source-signature prerequisite from #2347", so this fix is now load-bearing for A as well. That is a dependency #2335 does not declare (it lists A and E as independent). Either this lands before A, or A carries the same defect. I am recording it on the umbrella and sequencing E ahead of A accordingly.

Generated by Claude Code