fix(session): reconcile severed long turns against server state — no false 'failed to send' (#1376)

[AndriiPasternak31]

Summary

Fixes the Session-mode false-failure (#1376): a long-running turn shows "Failed to send message" while the backend execution is still running and ultimately succeeds. The Session turn endpoint is synchronous, so an intermediary with a read timeout shorter than the axios ceiling (notably an external Cloudflare ~100s edge timeout) can sever the socket mid-turn — axios sees a network error and the catch rendered an unconditional failure.

This is the stay-on-page / proxy-timeout variant of the previously-fixed #759 (return-navigation variant). Frontend-only fix; no backend or schema change.

What changed

• Reconcile before failing (stores/sessions.js): on any send error, sendMessage now reconciles against GET /sessions/{id} before declaring failure.
  • A new assistant beyond the pre-send baseline → success (checked first, because the assistant row persists inside the inflight sentinel — handing off there would let the poller rebaseline past the landed reply and false-fail).
  • turn_in_progress=true → hand off to the reattach poller (spinner stays, no error).
  • Neither → surface the real error; a pre-persistence drop restores the optimistic user row so the typed message stays visible.
  • A 429 concurrent-turn reject short-circuits before the reconcile (the sentinel reflects the OTHER turn).
• Wall-clock poll deadline (startPolling): replaces the MAX_ATTEMPTS=60 (~22 min) attempt count — unreliable under background-tab timer throttling — with a ~2h deadline covering the sentinel TTL. The deadline outcome routes through a neutral notice (noticeBySession), not a red error.
• UX (SessionPanel.vue): neutral status box (role="status"), aria-live on both error and notice so async outcomes are announced, and a "Still working…" spinner copy after ~75s kept alive across the poller handoff.
• E2E (e2e/session-tab.spec.js): T1–T8 stub the POST (sever / 502 / 429) and GET with server-truth fixtures — no real Claude turn.
• Docs: session-tab feature flow + Recent Updates index row.
• Security: /cso --diff daily report — clean, no findings.

Acceptance criteria

• A turn outliving an intermediary timeout never shows a false "failed to send" while running or succeeded.
• On a client-side error, sendMessage checks turn_in_progress / latest assistant and reattaches instead of unconditionally setting errorBySession.
• The poll path is bounded by a wall-clock deadline (≥ sentinel TTL), not an attempt count below the execution timeout.
• A genuine backend failure still surfaces a real error (no regression).
• Reverse-proxy read timeout: identified as an external Cloudflare edge timeout (operator config, outside this repo); the turn is now made resilient to the dropped socket rather than requiring a proxy change.

Closes #1376

🤖 Generated with Claude Code

[vybe]

Validated via /validate-pr: Closes #1376, e2e T1–T8 regression cases name the issue, session-tab flow doc updated, CSO --diff clean, security scan clean. Reconciles severed long turns against server state — no false 'failed to send'. LGTM.