fix(gateway): recover long correlated Claude results from the transcript

[Yehonal]

Long Claude replies were being lost over MCP even though the agent answered correctly with the right result markers. Short tasks worked, but every review-sized fresh task failed with result_uncorrelated or result_timeout.

Two causes:

• the terminal collector drops a reply that scrolls out of the pane, and gives up after timeoutSeconds while a long turn is still running;
• the transcript fallback used the public search path, which clips each message to 4096 characters, so the closing marker of any longer reply was cut off.

Changes

• Transcript result reader. New agent-session.sh --agent claude result <session> --correlation-id <id> [--wait <s>] (claude-transcript-result.mjs) returns the whole reply from the native transcript. It accepts only one assistant text block with exactly one marker pair; duplicates or an unterminated pair give exit 67, nothing found gives 124.
• Provider recovery. After result_uncorrelated or result_timeout the provider calls that reader and waits up to agentSessions.resultRecoverySeconds (0 to 14400, default 0 = check once). The task stays working while it waits. The search-based fallback is gone.
• 8 KiB truncation. agent-native-call.mjs wrote the result and called process.exit() right away, which cuts piped stdout at 8 KiB, so large native-session results came back truncated but marked completed. Both scripts now write synchronously. A new test with a 200 KB result fails on the old code.
• Locale. With LANG=C (how the gateway service runs today) tmux and grep mangle the ❯/· glyphs used by readiness and trust-dialog matching. _mesh-tmux.sh now switches to C.UTF-8 when the locale isn't UTF-8. This was also why launch-options-test.sh failed locally but not in CI.
• Fresh follow-ups. The new target is killed when writer ownership can't be proven (nothing was sent yet). --resume A --fork-session no longer counts as a writer of A. Added a test for fresh_session_unsupported raised at execution time.

Tests

• npm run verify passes on a LANG=C host, including the full tmux-bridge harness (75/75) and the gateway suite (254/254).
• The new reader was checked read-only against four real failed fresh sessions: it returns the complete replies (5.7 to 26 KB).

Rollout note

To make long reviews complete, the live config needs resultRecoverySeconds (e.g. 3600) after deploy. With the default of 0, only replies already in the transcript are recovered.