feat(pi): carry thinking_delta as ChatDelta.reasoning so retained_events keeps the model's reasoning

[drewstone]

Problem

The pi backend drops every thinking_* event from pi --print --mode json (src/backends/pi.ts, pinned by tests/pi-backend.test.ts), and ChatDelta has no reasoning field, so the model's reasoning never reaches the retained event log. The reasoning does arrive: in a running fleet (pi 0.84.2, provider tangle-router, model openrouter/stealth/ox-alpha, --thinking high), pi's own session files hold 556 thinking blocks across 19 of 22 sessions — 1.89 M reasoning characters against 108 K visible-text characters — while retained_events for the same sessions holds 0 rows with any reasoning key (23,110 rows inspected). The bridge's inference proxy is not the cause; it forwards the request body byte-for-byte and pipes the SSE response through identity and redaction transforms only.

Change

1. ChatDelta.reasoning?: string (src/backends/types.ts): incremental reasoning text, never assistant content.
2. src/backends/pi.ts: thinking_delta no longer hits the drop path. Deltas buffer and flush as one { reasoning } ChatDelta; thinking_start / thinking_end remain dropped as boundary markers. A reasoning block ends when the next non-reasoning delta arrives.
3. src/streaming/sse.ts: the SSE writer maps reasoning to choices[0].delta.reasoning — the field OpenRouter and pi already use. collectNonStreaming output is unchanged; reasoning never enters message.content.
4. src/runs/replay-log.ts: replay retention accounting counts reasoning bytes. src/routes/chat-completions.ts: the estimated-usage fallback counts reasoning characters as output.
5. No store, route, run-registry, or replay-log structural change: appendRetainedDelta already persists the delta verbatim under {"type":"raw","backend":"cli-bridge.chat","event":…}.

Coalescing rule and cost

Pi emits one thinking_delta per provider chunk (a few characters), and reasoning outweighs visible text ~17:1 by characters, so per-chunk rows would multiply the retained-event row rate ~18x (about 400,000 extra rows for the 31 runs measured). The backend buffers reasoning and flushes at 256 characters, at 250 ms of buffer age, or before any non-thinking yield — ordering against content / tool_calls / usage stays exact, a live reader sees reasoning within a quarter second, and the row count divides by roughly 50.

Acceptance

• A fleet run on pi with reasoningEffort: high produces retained_events rows whose inner event has reasoning, interleaved before the content rows of the same assistant message.
• collectNonStreaming output is unchanged (reasoning never enters message.content).
• The existing test's content deltas are unchanged.

Research: tangle-network/discovery-lab/docs/research/reasoning-deltas.md (on master).