[jaredLunde]

This PR fixes the two LOW audit findings on #156. Each fix has a test that fails without it.

1. The restore window never waits on a session that will not restore

Problem: a listed restoring session could hold the 503 window for its full 10 minutes in four cases:

• it failed to start (serve_ws only logged);
• its body ended with an error;
• it got no events hub because its MCP catalog is now empty;
• it panicked on every start.

Fix: mcp_events::restored(session_id) (now public) is called on each terminal path:

• a failed boot pin, for the events session and for each runtime session;
• serve_session_body ending with an error;
• serve_session starting with no events hub;
• every panic. A panicked incarnation reserves nothing more. Tokens it already reserved stay held, and a restart reserves its own again.

Tests (mcp_events_receiver):

• a_restoring_session_with_no_events_hub_closes_the_restore_window: the session's server is no longer configured on restart.
• a_restoring_session_that_fails_to_start_closes_the_restore_window: its transcript no longer opens.

Each asserts that a token nothing holds is 410 within 10 s. With the terminal calls reverted, both stay 503 and fail.

2. The release binary is the authority on "no test seam ships"

• New script: crates/agent/scripts/scan-release-seams.sh greps each given artifact for BEYOND_AI_AGENT_TEST_ and fails on any hit, listing the names it found. It checks what the compiler kept, so source tricks don't matter: a concat!-assembled name, a // inside a string literal, anything a text heuristic misreads.
• CI: the build (release) job now runs mise check:release-seams right after mise build:rs:release, over every top-level release binary. It reuses the job's own artifacts, so there is no extra build and it adds seconds. The agent test shards are untouched.
• Source checker: tests/release_seams.rs stays as the fast early warning and now says the binary scan is the authority.
• Scanner test: release_seams::the_release_artifact_scan_fails_on_any_seam_name. The scanner must fail on this crate's debug build (which carries the seams by design), must fail on a concat!-assembled name, and must pass a clean file.
• Local check: the scan passes on a release build and fails on the debug build.

Verification

• clippy (-D warnings, all targets), cargo fmt and dprint check are clean.
• lib: 1600 passed.
• These targets pass: mcp_events_receiver, mcp_events_nested, mcp_events_lifecycle, mcp_events_durability, release_seams, serve_service_session_end, serve_session_lifecycle, serve_harness_deadlines.
• ARCHITECTURE is updated for both changes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01JimHGjsfk2Ktm5GxyZJKKk

[jaredLunde]

Superseded by #161 (the release-binary seam scan, split out). The restore-window half is dropped: it covers a narrow MCP Events webhook edge case, and main's current behaviour (a retrying server may get 503 for up to 10 min after a restart) is acceptable.