[golimpio]

What

This fixes a latency regression from #648 (caller completeness). For every recorded call of a seed's name, the census asked whether the call's file dot-imports a package. SQLite answered that through the import-kind index, scanning every import node in the workspace once per call site. For a common method name, such as Execute (962 sites against 11k imports), that took seconds.

This is how arm R (PLAN-543, real open plumb bugs) found it:

• In 3 of the 5 directed-use packs, the affected-tests section said only "the 2s graph deadline passed, so the estimate is partial".
• So the named-tests run-first line from feat(context): name the tests a call edge reaches, as a ready run_task filter #647 reached the agent in 1 of 5 packs.

The census runs after the walk and before the test estimate, and it was using up the shared clock.

Fix: the set of dot-importing files is read once per census, by one indexed query that takes about 9 ms, and looked up per site in Go. The census claims don't change.

Measured on arm R's actual requests

These are the exact context_for_task arguments from the D transcripts, replayed in-process on the same index (plumb dfa964f, 2,016 files), warm, 3 reps each:

pack	before	after	run-first line
R05 (#601)	7.5–9.6 s	0.36 s	missing → present
R08 (#600)	6.4–10.1 s	0.23 s	missing → present
R03 (#599)	1.2 s	0.25 s	present
R01 (#642, control)	0.15 s	0.23 s	present (kept)

A CPU profile before the fix put 91% in Store.OpenCallSites.

Tests

• TestOpenCallSites_CostDoesNotScaleWithImportsPerSite: 2,000 calls of one name over 20,000 imports must finish within 1.5 s. The fixed query takes about 0.29 s.
  • Positive control: with the per-site probe restored by hand, it fails at 14.9 s.
  • The bound is about 50× wider than the fixed query's time, so it measures the query's shape, not the machine.
• TestOpenCallSites unchanged: the dot-import flag still drives the bare-call rule, and feat(context_for_task): say when a seed's callers are complete (per-call resolver outcome, schema 8) #648's mutation run killed the "flag lost" mutant.
• ./internal/topology/... and ./internal/tools both pass.

Why #648 missed it

#648 was verified for correctness (type-checker oracle, mutation tests, container verify-full) but not for latency on the real repository. The gate-v1 latency harness had run on the collector before the census existed. Lesson recorded: a change to context_for_task's collection path gets a replay-latency check against real requests before merge.

🤖 Generated with Claude Code

[atlas-from-plumb]

Approved under the owner's standing delegation. The bar is met on head 73096a1:

• Independent review (scarlet-badger): PASS. Mutation run in its own worktree: DotImport wiring forced false, KILLED; set query on the wrong name, KILLED. Timing test checked under -race at 285 ms against the 1.5 s bound.
• Positive control: with the per-site probe restored, the cost test fails at 14.9 s.
• CI 9/9 on 73096a1.
• make verify-full-container: exit 0, 61 ok, at most 1 daemon, Revision 73096a1.
• Up to date with main 995915c.