perf(server): adopt the upstream server performance pass

[rynfar]

Five upstream perf commits, in order:

• #8988 stop loading full tool-call output from the DB on every refresh
• #9000 bound snapshot activity payload memory
• #9032 skip full-message reads while streaming
• #8187 cut idle CPU use and stop provider event leaks
• #8368 cut chatty tool-update frames by 90%

#9024 was already on pylon, so its cherry-pick was correctly a no-op.

Carve-outs

ProjectionSnapshotQuery. #9000 extracts the pinned-activity CTE into
pinnedThreadActivityIdsCte. Pylon pins one more thing than upstream — the latest
turn's authoritative plan, via latest_plan_activity — and its tests assert 504
activities where upstream asserts 503. Both are preserved. Upstream also
reintroduced a plainer ProjectionThreadSessionDbRowSchema alongside Pylon's,
which carries restored and pendingTurnRequestAmbiguous; the duplicate is
dropped and Pylon's kept.

ProviderRuntimeIngestion. #8187 adds an early return for non-assistant
content deltas and drops the unconditional getPendingTurnStartByThreadId read in
favour of an event-gated one. Both are taken. Pylon's rate-limit handling and its
turn.completed capacity refresh — the composer capacity gauge — sit in the same
hunk and are kept, as does let thread, which Pylon reassigns further down.

Two regressions this surfaced, both fixed in the last commit

Pylon's pending-input count silently went to zero. #8988 narrows the
thread-shell summary read to a query whose SQL filters to the three user-input.*
kinds. Pylon's Prime adapter emits generic interaction.* activities for the same
count, and shouldRefreshThreadShellSummary already wakes on them — with a
comment saying so. So the refresh fired, read no rows, and wrote zero. Nothing
typechecked wrong; ProjectionPipeline.test.ts caught it.

Prime protocol logging. #8187 puts raw ACP protocol logging behind a new
verboseProtocolLogging flag that no production adapter sets — upstream's own
Grok and Cursor adapters included, and the flag appears set only in upstream's
unit test. Pylon's Prime thought-chunk test asserted an outgoing protocol frame
reached the native log. That frame no longer exists for any ACP provider, so the
test now asserts the stronger property instead: no protocol frames at all, which
means a thought chunk has no path to disk.

Verification

t3 typecheck clean, lint clean, 3890 server tests passing (2 were failing
before the fixes above; both bisected to the commit that caused them).

Model: Claude Opus 5 in Claude Code.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.7 KiB	13.5 KiB	−210 B (−1.5%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	−7 B (−0.1%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.8 KiB	6.6 KiB	−203 B (−2.9%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	58.7 KiB	57.1 KiB	−1.5 KiB (−2.6%)	66.4 KiB	✅
Codex	Live turn messages	11	9	−2 (−18.2%)	21	✅
Claude	Total thread wire	13.7 KiB	13.4 KiB	−361 B (−2.6%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	0 B (0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.8 KiB	6.5 KiB	−361 B (−5.2%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	59.5 KiB	57.2 KiB	−2.4 KiB (−4.0%)	66.4 KiB	✅
Claude	Live turn messages	11	8	−3 (−27.3%)	21	✅

Baseline: 4564313 · PR result: 02b9a45 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 109.5 KiB
• Claude decoded thread snapshot: 110.2 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.