pr #28930 runtime verification
main: 71646e031a6dc9db1309b3b9dff597f872cf2fd3
pr head: a32e316cd1e1150d21a4e218c3d883e9af3d1e32
client: repository Playwright 1.60.0 + /usr/bin/google-chrome, Linux
provider: real Codex gpt-6.1-sol through local proxy
data: two conversations created with the web client and completed real provider turns on main; isolated state copied after stopping main, preserving identical messages for PR comparison. A third real provider turn completed on PR head. No mocks, demo provider, SQL fixture writes, or direct matcher calls.

$ sidebar search -> orchestration.searchThreads -> SQLite -> messageMatcher -> client results
mode: PASS identical matches, snippets, quality, source, hits, message IDs and order (2 threads)
cafe: PASS identical matches, snippets, quality, source, hits, message IDs and order (2 threads)
resume: PASS identical matches, snippets, quality, source, hits, message IDs and order (2 threads)
same same: PASS identical matches, snippets, quality, source, hits, message IDs and order (1 threads)
vp run dev: PASS identical matches, snippets, quality, source, hits, message IDs and order (2 threads)
gannet's: PASS identical matches, snippets, quality, source, hits, message IDs and order (1 threads)
部署: PASS identical matches, snippets, quality, source, hits, message IDs and order (1 threads)
absentsearchzqx: PASS identical matches, snippets, quality, source, hits, message IDs and order (0 threads)
assistantneedle: PASS identical matches, snippets, quality, source, hits, message IDs and order (1 threads)

$ live server inspector: non-pausing conditional breakpoint at findMessageToken loop body during sidebar search
same 32 KB user message, 6,400 repeated mode words plus prefix/whole-word controls
main quoted mode search: 6,406 iterations
PR single-token mode search: 2 iterations (model prefix, first whole-word mode)
This counts actual matching work during the live search, not whole-app latency. No end-to-end speedup or allocation multiplier is claimed.

PR fresh turn response: saffronlighthouse İ Café Café
sidebar search saffronlighthouse: source=assistant, quality=1, one thread, exact generated snippet
command palette Ctrl+K search saffronlighthouse: same thread under In messages, Agent snippet
click result: opens correct conversation, find-in-thread highlights generated word, 1/1

remote checks via blacksmith-check:
vp test run packages/shared/src/paletteSearch.test.ts apps/server/src/orchestration-v2/ThreadSearch.test.ts: 60/60 passed
vp run check:changed fa31b66bbb2ae5234205a8628a78e038d917202f: format, lint, knip, affected typechecks (33 packages), tests (35/35) passed
initial check:changed could not find merge base in runner shallow snapshot; fetched PR history with depth 2, then passed.

limitations: Linux web only. Native desktop/mobile, other providers, relay/tunnel, production-sized history, and the author's benchmark multipliers were not verified.
State copied from newer main has migration 77, unknown to PR build (76); server warned and started. No migrations were altered. Functional searches and a fresh PR provider turn succeeded.
No repository source edits, commits, pushes, labels, merges or branch history rewrites.