The home page and repo page still stall for about a second on a cold authenticated request #7940

Pushed, but I did not get / under 50 ms. The cold case is still about 390 ms on my test data, and I couldn't check the real 2.46 s figure here.

The two requests slow down in two ways:

• Telemetry kept the caches cold. The home's KPI, vitals and waiting caches are valid only while the forge-wide write clock is unmoved. Every signed-in visit writes a few rows of its own (route timings a second later, usage counts, flag reads, the session's last-seen time, the home visit, page vitals, machine and memory samples). The next visitor therefore always found the caches cold and recomputed everything.
• Several cold reads scanned whole tables. The open counts for each repository ran about 16 times per page, plus the open-pull vitals, the "ready asks" tally and the parked-by-note check.

Changes

• quietly(db, write) in src/core/write_clock.ts. It leaves the rows a telemetry write changes out of the stamp. It now wraps the writers listed above. Any other write still moves the clock. The caveat is in docs/CACHING.md: only wrap tables no cached page is built from.
• Migration 20261102120000-cold-home-indexes (pinned, additive):

- issue(repo_id,state,sealed,kind) covers the per-repo open counts, on both the home and repo pages. - A partial index for open pulls, issue(created_at) for the vitals time windows, and comment(issue_id,created_at,id).

• Rewritten queries:

- The ready-asks query in flow/pipeline.ts now walks autonomous repos and then their open issues, not every issue. In the micro-benchmark that went from 28 ms to 7 ms. - The open-pull vitals query in flow/vitals.ts now filters on kind pull, so the new partial index applies. - breakerParked in crew/breaker.ts now reads each issue's newest comment through the new index instead of windowing over all comments. It went from 38 ms to 17 ms in the micro-benchmark.

Measured

• Data: a local fixture of about 12.8k issues, 4.8k pulls and 13.6k runs.
• Preview baseline: the previews answer /healthz in about 1 s through my proxy, so I couldn't separate server time from network there. The server-timing header says the app takes under 20 ms.
• Cold /: about 616 ms before and about 385 ms after.
• Cold repo page: about 40 ms before and 17–22 ms after.
• Repeat visits: those no longer go cold from the visit's own writes. test/write-clock-quiet.test.ts covers this, and it fails without the change.

Left out

• The cold home is still dominated by three reads that grow with the total number of asks and pulls: askedCounts (about 90 ms), pullsByAsk (about 50 ms) and the pulls-in-flight vitals (about 60 ms). Getting those under budget needs a different structure, such as incremental counters, so I did not rewrite them here.
• Real crew writes (events, runs) still invalidate the caches, as they should.
• I haven't checked the preview of this pull yet.

Tests, tsc and scripts/gate.ts pass on what I ran. scripts/fast.ts named 1273 test files, so I ran the schema, home, flow, snappy, route-timing, usage-count, flags and vitals suites instead.

Closes #7896

Pushed, but I did not get / under 50 ms. The cold case is still about 390 ms on my test data, and I couldn't check the real 2.46 s figure here.

The two requests slow down in two ways:

• Telemetry kept the caches cold. The home's KPI, vitals and waiting caches are valid only while the forge-wide write clock is unmoved. Every signed-in visit writes a few rows of its own (route timings a second later, usage counts, flag reads, the session's last-seen time, the home visit, page vitals, machine and memory samples). The next visitor therefore always found the caches cold and recomputed everything.
• Several cold reads scanned whole tables. The open counts for each repository ran about 16 times per page, plus the open-pull vitals, the "ready asks" tally and the parked-by-note check.

Changes

• quietly(db, write) in src/core/write_clock.ts. It leaves the rows a telemetry write changes out of the stamp. It now wraps the writers listed above. Any other write still moves the clock. The caveat is in docs/CACHING.md: only wrap tables no cached page is built from.
• Migration 20261102120000-cold-home-indexes (pinned, additive):

- issue(repo_id,state,sealed,kind) covers the per-repo open counts, on both the home and repo pages. - A partial index for open pulls, issue(created_at) for the vitals time windows, and comment(issue_id,created_at,id).

• Rewritten queries:

- The ready-asks query in flow/pipeline.ts now walks autonomous repos and then their open issues, not every issue. In the micro-benchmark that went from 28 ms to 7 ms. - The open-pull vitals query in flow/vitals.ts now filters on kind pull, so the new partial index applies. - breakerParked in crew/breaker.ts now reads each issue's newest comment through the new index instead of windowing over all comments. It went from 38 ms to 17 ms in the micro-benchmark.

Measured

• Data: a local fixture of about 12.8k issues, 4.8k pulls and 13.6k runs.
• Preview baseline: the previews answer /healthz in about 1 s through my proxy, so I couldn't separate server time from network there. The server-timing header says the app takes under 20 ms.
• Cold /: about 616 ms before and about 385 ms after.
• Cold repo page: about 40 ms before and 17–22 ms after.
• Repeat visits: those no longer go cold from the visit's own writes. test/write-clock-quiet.test.ts covers this, and it fails without the change.

Left out

• The cold home is still dominated by three reads that grow with the total number of asks and pulls: askedCounts (about 90 ms), pullsByAsk (about 50 ms) and the pulls-in-flight vitals (about 60 ms). Getting those under budget needs a different structure, such as incremental counters, so I did not rewrite them here.
• Real crew writes (events, runs) still invalidate the caches, as they should.
• I haven't checked the preview of this pull yet.

Tests, tsc and scripts/gate.ts pass on what I ran. scripts/fast.ts named 1273 test files, so I ran the schema, home, flow, snappy, route-timing, usage-count, flags and vitals suites instead.

This pull only partly fixes the cold authenticated home and repo page. I did not measure cold / or the repo page on the preview with real-size data. The last figure I have is the builder's cold / at about 385 ms, so the <50 ms target is not shown to be met. I split the remaining reads (askedCounts, pullsByAsk, and the pulls-in-flight vitals) into #7953 and did not link it to the umbrella issue.

What this pull does, as before: telemetry writes no longer move the write clock, so the home caches stay warm. The ready-asks, parked-notes and open-pull reads now use indexes.

I ran the two new test files, tsc and the gate; all pass. I did not run scripts/design-lint.ts, because no src/web markup changed.

Changes in response to review:

• Per-writer tests (item 2): test/write-clock-quiet.test.ts now has one test for each of eight telemetry writers: route timing, usage count, flag read, session last-seen, home visit, page vital, memory sample and machine sample. Each asserts the clock stamp is unchanged. With the wrap in quietly disabled, all eight fail. I disabled it for all writers at once, not one at a time.
• Table guard (item 5): the same tests dump every table before and after each writer and assert it changed only an allowlisted telemetry table. A loud write inside quietly fails that check. A separate test asserts a loud write still moves the clock.
• No sleep (item 4): Bun.sleep(1300) is gone. The cache-hit test fires the timing flush and a page vital directly, then asserts the second home read is a hit.
• Query equivalence (item 3): the new test/cold-home-queries.test.ts compares the rewritten queries with the old SQL on rows built to tell them apart.

- breakerParked: one issue parked by its newest note, one whose parked note is no longer the newest, and one closed issue. The result equals the old window-function result. - Open-pull vitals: an open non-pull issue is not counted, and the count equals the old query's. - Ready asks: the tally equals a hand-counted version of the old join, and drops to 0 when autonomy is off.

• home_visit (item 6): the cached KPI, vitals, waiting and panel values never read it. lastVisit is read in home.ts and passed to the "since your last visit" block, which is rendered uncached for each request. A test asserts home_cache.ts never mentions home_visit or lastVisit. Keeping the visit write quiet is therefore correct.