fix(server): reveal buffered prose before tool activity

[rynfar]

When a provider emits an assistant paragraph and then starts a tool before completing the assistant item, Pylon can show the tool activity while the prose stays buffered. This makes the stream appear to stop until a later paragraph or item completion. Flush that turn's buffered assistant text immediately before appending a tool.started activity, preserving the current Markdown-aware buffering and normal delivery rate for prose outside tool boundaries.

A regression drives Claude-shaped runtime events through the real ingestion harness: one unfinished paragraph, a tool start, then more assistant text and completion. It asserts the first prose is visible before the tool activity and the final message is complete without duplication. This is a server projection fix shared by web, desktop, and mobile; it does not change provider protocol handling or the user setting for legacy token streaming.

Validation: all 124 ProviderRuntimeIngestion tests, server typecheck, scoped lint and format, and git diff --check passed. One pre-existing unused-import lint warning remains in the touched server file. No live Claude turn was run, so this addresses the identified buffer/tool ordering path rather than claiming every cause of a missing reply is resolved.

---

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

[vercel]

Deployment failed for project pylon-marketing with the following error:

Resource is limited - try again in 24 hours (more than 100, code: "api-deployments-free-per-day").

Learn More: https://vercel.com/pylon-code?upgradeToPro=build-rate-limit

[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	13.9 KiB	13.9 KiB	+13 B (+0.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	7.2 KiB	7.2 KiB	+4 B (+0.1%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.7 KiB	6.7 KiB	+9 B (+0.1%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	58.0 KiB	58.0 KiB	+44 B (+0.1%)	66.4 KiB	✅
Codex	Live turn messages	9	10	+1 (+11.1%)	21	✅
Claude	Total thread wire	14.0 KiB	14.0 KiB	+23 B (+0.2%)	15.1 KiB	✅
Claude	Thread snapshot wire	7.2 KiB	7.2 KiB	+4 B (+0.1%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.8 KiB	6.8 KiB	+19 B (+0.3%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	58.9 KiB	58.9 KiB	0 B (0.0%)	66.4 KiB	✅
Claude	Live turn messages	10	10	0 (0.0%)	21	✅

Baseline: afd6c11 · PR result: d6c0932 · 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: 115.6 KiB
• Claude decoded thread snapshot: 116.3 KiB

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

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
pylon-marketing	Ready	Preview	Sep 26, 2026 6:48am UTC

[rynfar]

Refreshed onto current pylon at d6c0932fce with a clean merge. The 124 focused ingestion tests, server typecheck, scoped lint/format, and diff check pass on this head. The prose-before-tool regression still exercises the real ingestion harness. Waiting for exact-head CI before merging; no live Claude turn or device run is claimed.