feat(cesar): recap reflects real outcome, not tool noise — + fork-men…

[cukas]

…u fix

Recap redesign — the end-of-turn summary stopped over-signalling failure:

• Headline ✓/⚠ derives from the TURN outcome (did the engine respond), never from intermediate tool exit codes, so a shipped turn no longer reads as ⚠.
• Per-command ✓/✘ list dropped. Only UNRECOVERED CONSEQUENTIAL failures get a red ✗ line, via a typed classifyConsequence (commit/push/build/test/… vs plumbing, fail-open on unknowns, segment-aware so cd && git commit is a commit). Recovered/plumbing failures fold into a dim "N non-fatal" count that by construction can never hide a real problem.
• The benign "paused mid-task — auto-continuing" nudge is marked transient at its source and excluded from the recap (still shown live as it happens).
• /compact now shows "⤵ context compacted · session continues".
• changes:/verified: success rows kept in parity across the live and scrollback renderers; the layout row estimator tracks the new rows.

Fork-menu fix — the "pick one" picker fired on the wrong list:

• New, unit-tested extractAdjacentForkOptions: a picker fires only when the numbered/lettered options are a contiguous block directly above the closing question. A numbered DESCRIPTION followed by prose and an open question is no longer mis-rendered as a picker; option labels truncate on word boundaries.

Hardening (agon review panel):

• OutputEvent cesar-recap variant and warning.transient are now typed (no as any); verification reflects the LAST attempt per label (an early failed build/test that a retry fixed reads ✓); the emit gate no longer treats a lone plumbing command (git status) as a finding.
• _noBriefNudged: module Set → WeakMap keyed by the session object (no cross-session leak); dispatchToolCall duration map → per-key FIFO so concurrent same-name tools can't corrupt each other's durations; KERN_VERSION fallback 3.5.7 → 4.0.0.

+30 unit tests; full suite green.

⚔️ Forged by Agon