fix(chat): keep the full agent trace visible in history and never end a turn without a summary

[ABB65]

Problem

Follow-up to #159 (2/2). Persisted transcripts hid every intermediate iteration: only the final assistant row was `internal=false`, so a turn that burned all 8 iterations on tool calls showed the literal `[tool calls]` placeholder as its entire answer on reload (verified on staging, conversation `8ad2bba3…`). The graceful-close wrap could also return zero content without erroring — no error in Railway logs, no summary in the trace.

Changes

• buildTraceRows: every assistant iteration row is now visible; `tool_result` payload rows (the big truncated JSON blobs) stay internal. Token/model columns still land only on the final row. No migration needed — RLS policy 009 (`internal = false`) exposes the rows as-is.
• Graceful close hardening:
  • The tools-disabled wrap now carries an explicit summarize instruction appended after the final `tool_result` blocks. The message is replaced, not mutated — the trace holds the same content array reference, so the instruction is never persisted or replayed on resume.
  • If the wrap still returns no text, a deterministic fallback summary built from the executed tool names streams to the client and lands as the final trace iteration — a transcript can never end tool-only again.
• Messages endpoint: projection gains `turn_id`; limit 100 → 300 (heavy turns now surface up to ~10 visible rows instead of 2; the oldest-N-under-ascending-order wart is noted in a comment, tracked separately).
• loadConversation: consecutive same-turn assistant rows fold into one message whose segments concatenate chronologically — reload now mirrors the live stream. Legacy rows (distinct turn ids) keep one message per row.

Documented side effect: the ee Conversation API `/history.get` uses the same default visibility filter, so intermediate assistant rows now appear there too (additive). Existing conversations are not backfilled — old turns keep their previous single-row rendering.

Test

• `pnpm test` — 130 files / 1078 tests green
• `pnpm test:contract` — 105 tests green against local `postgres:16` (fixtures updated to mirror the new writer)
• Engine regression: wrap-instruction shape + trace purity + zero-block fallback covered
• `pnpm lint` + `pnpm typecheck` clean

🤖 Generated with Claude Code