[atlas-from-plumb]

What

After #663 a pack asks the sensitive-path decision once per distinct file, which is 100 to 250 files. Each ask still resolves the file's project with workspacePool.Detect, which globs for markers on disk. On the live daemon the slowest arm R pack (R05) still runs at 1.6–1.9 s of its 2 s graph deadline. In a baseline taken today on fa1fd85, a second pack (R08) was cut in 2 of 5 reps, losing its affected tests.

The files of one directory resolve to the same project, so within one call this PR detects each directory once.

• tools.CallMemo: a small lookup memo carried on the call's context. Collect attaches a fresh one on every call; a nil memo runs every lookup.
• historyRootFor (cli): detects through the memo. Outside a context_for_task call, which covers every write, there is no memo, so detection runs on every ask exactly as before.
• changeSensitive stays the one decision. History and the response gate still share it.
• workspaceFromArgs is split into detectStartDir and detectedRoot, with identical behaviour; its existing tests pin it.

Because the memo lives on the call's context, it can't outlive a call, and a marker created between two calls is seen by the second.

Tests

• TestCallMemo_RemembersPerKeyAndNilRemembersNothing: one lookup per key, and a nil memo looks up on every ask.
• TestContextForTask_SensitiveAsksShareOneMemoPerCall: every ask in one call carries the same memo, and the next call gets a fresh one.
• TestHistoryRootFor_DetectsEachDirectoryOncePerContextCall:
  • a marker created part-way through a call is not detected again in that call, which proves the second ask reused the first detection;
  • the next call sees the marker;
  • so does an ask outside a call.
• TestChangeSensitive_CallMemoKeepsEveryVerdict: each case is asked twice in a call and must give the same verdict as without the memo. Cases cover an ordinary file, a default glob, a project's own glob, the same name where no glob applies, and a move out of another project's file that is sensitive only by that project's glob. Controls require both verdicts to appear and the cross-project case to be sensitive.

Mutation (plumb mutation_test): all 7 valid mutants are killed:

• Collect attaches no memo;
• Once never stores;
• a nil memo returns empty;
• historyRootFor ignores the memo;
• the memo uses one constant key, which leaks a verdict across directories;
• detectStartDir always takes the parent (killed by the existing workspaceFromArgs and attribution tests);
• historyRootFor bypasses the memo by passing a context with none.

Measure

A live replay on the installed build after merge, with per-pack latency, to show R05 well under 1.5 s and no cuts under sample. Baseline on fa1fd85 (5 reps): R01 0.36–0.42 s, R03 0.64–0.75 s, R05 1.58–1.66 s, R08 1.05–2.09 s with 2 of 5 cut.

Verification

make verify-full-container: result to follow in a comment. Review: a peer agent (warm-cobra), who filed PLAN-548.

🤖 Generated with Claude Code

[golimpio]

Approved under the delegated merge bar, on head bdff688:

• independent review by a peer agent (warm-cobra): PASS on 1204833; its NIT is folded in as bdff688, a comment-only change;
• CI 9/9 green on this head;
• make verify-full-container exited 0 on this head;
• plumb mutation_test: 7/7 valid mutants killed.