feat(sessions): configurable per-session exclusivity with queue-instead-of-refuse (#101279)

[salch-cred]

What

Configurable per-session exclusivity for shared-brain deployments (#101279): a new gateway.per_session_exclusive config switch (default true — behavior unchanged) that, when set to false, turns a SESSION_NOT_OWNED refusal into a bounded wait for the current owner to finish, then re-acquire — the queue semantics the messaging gateway already gives Telegram — instead of pushing a mid-day failure onto the second surface.

This implements the issue's proposed Option 1 (config switch, default on) + Option 2 (queue instead of refuse) as one coherent design: the switch decides whether the desktop/CLI submit paths queue or refuse.

Changes

• hermes_cli/active_sessions.py
  • per_session_exclusive(config) — resolves gateway.per_session_exclusive (top-level fallback, bool/"true"/"false" strings accepted, invalid values warn once and keep the safe default True).
  • wait_for_session_ownership(...) — bounded poll of the live-lease registry until no other lease holds the session (or timeout/abort). Never raises: an unreadable registry returns False so the caller falls back to the original refusal. on_wait fires exactly once for a status line ("Another Hermes process is using this session; waiting...").
• tui_gateway/session_lifecycle.py — _ensure_active_session_slot: on a SESSION_NOT_OWNED refusal, when the operator opted out, wait for the owner and re-acquire (_maybe_queue_for_session_ownership). All other refusal reasons (registry unavailable, capacity) are untouched, and the default path is a single dict check as before.
• cli.py _claim_active_session — same queue-instead-of-refuse on the CLI surface, with a yellow hourglass notice.
• tui_gateway/methods_voice.py gateway.capabilities — advertises per_session_exclusive_queue_mode (config-dependent, separate from the build-enforced per_session_exclusive_submit) so clients can surface "queued" instead of "refused" only when it is actually true.

What does NOT change

• Default deployments: exclusivity stays exactly as enforced today (the existing tests/test_active_session_exclusivity.py suite passes untouched).
• SESSION_COORDINATION_UNAVAILABLE (registry unreadable) still fails closed — queueing is never used to smuggle a second writer through an unknown-ownership state.
• MAX_CONCURRENT_SESSIONS capacity policy is untouched.
• The wait is bounded (default 1800s, matching LEASE_WAIT_SECONDS); a timeout falls back to the original refusal.
• PER_SESSION_EXCLUSIVE_SUBMIT remains a module constant: the build-level enforcement is unchanged; the config only governs how a surfaced refusal is handled.

Tests

tests/test_per_session_exclusive_opt_out.py (13 new):

• resolution: default true; gateway false/true; top-level fallback; string coercion; invalid values warn + keep default; attribute-style config objects
• exclusivity still enforced when opted in
• wait semantics: free session returns immediately; held session blocks until release then re-acquire succeeds; bounded timeout returns False and the refusal still applies; on_wait fires exactly once (not per poll); other sessions' leases are not waited on

Regression: tests/test_active_session_exclusivity.py + tests/hermes_cli/test_active_sessions.py (29) pass unchanged; tests/tui_gateway/test_protocol.py (68) and the cross-process ownership suites (10) pass. ruff check clean.

Closes #101279

[alt-glitch]

This was generated by AI during triage.

Related: implements #101279 (Options 1+2). Note this branch also carries the unrelated agent/turn_context.py row-addressed api_content backfill and its test from #103721 (same author) — consider rebasing so this PR only contains the per-session-exclusivity change. Adjacent lease fixes in flight: #103713, #103737.