PR #29310 real-environment verification
main: a03648901640216a154d387cfb2cdec268e6527e (github/main)
PR:   4971ecfb4d8c469cfc6d288333394ded006d3690 (exact head)
Linux Boat VM; repository Playwright 1.60.0; /usr/bin/google-chrome.
One task dev server at a time, task-local homes, background processes.

Created project /home/user/t3code and thread through the web app.
Selected Codex gpt-6.1-sol in the composer and submitted a real turn.
Provider completed with markdown prose, a code block, and a bullet.
Stopped main, copied its isolated dev state, started the exact PR head.

Actual Find UI queries, completed result counts on both revisions:
  amber orchard    5
  violet harbor    4
  cobalt lantern   2
  absent sapphire  0
  amber orchard    5

Unmodified server.trace.ndjson, successful ThreadFind.load spans:
  main: 4 payload-batch loads for the 4 distinct queries
  PR:   1 payload-batch load for the same queries and conversation
  returning to amber orchard preserves the existing per-query cache
No user-perceived latency improvement is claimed for this small thread.

PR neighboring behavior:
  previous match wraps 1/5 -> 5/5; next wraps back to 1/5
  Enter and Shift+Enter navigate matches
  prose, user prompt, code block and bullet matches remain searchable
  absent query completes at 0/0
  searched fresh zephyr: 0/0
  closed Find, submitted a second real Codex gpt-6.1-sol turn
  provider completed: fresh zephyr confirms the second provider turn.
  searched fresh zephyr again: 1/2, next selects 2/2
  the previously cached zero-result query includes the new prompt and reply
  no browser page errors observed

Remote checks on task-owned Blacksmith testbox:
  vp test run apps/server/src/orchestration-v2/ThreadFind.test.ts
    20/20 passed
  vp run check:changed 47f2c8d29e981c7723467efe137ee19e91df924e
    format, lint, knip, GraphQL, affected typechecks: passed
    adjacent and cross-file tests: 21/21 passed
  Explicit base used because the worker source snapshot lacked main history.
  cafe pr status: 31 successful, 0 failed, 0 pending checks at this head.

Not exercised in real clients: native mobile, packaged desktop,
remote/relay/tunnel connections, other providers, large-thread memory limits,
concurrent scans, fork ancestry, rollback, cwd/skill-label invalidation.
Focused existing tests cover cache budget and several invalidation cases;
they do not replace real-client verification of those paths.
No source edits, commits or pushes were needed.