fix(herdr): add blocked-transition confirm delay to suppress transient OpenCode false alarms

[burakgizlice]

Problem

Herdr reports agent-status changes via its transition event stream. When a pane
transitions to blocked, firstmate immediately escalates it as a stale wake,
with the message: stale: <window> (herdr: agent blocked - waiting on human, escalated immediately, not via wedge timer).

The OpenCode TUI briefly shows a "waiting for input" state between tool calls
as the agent cycles through its working loop. Herdr correctly reports these
transitions as blocked -> working -> blocked -> working on every tool
call. The existing per-pane dedupe marker prevents re-escalating the same
blocked edge, but each working transition clears the marker, so the next
blocked edge is treated as a fresh actionable event -- producing a firehose
of stale wake notifications every few seconds.

This is especially visible during:

• Normal agent working cycles (transient prompts between tool calls)
• Long-running commands like sleep 600 && no-mistakes axi status (not a
  wedge, just the agent waiting for itself)

Solution

Add a re-check step in fm_backend_herdr_apply_transition (the herdr
backend function that routes transition events through the shared policy
table). When a blocked transition arrives:

1. Wait 3 seconds (configurable via FM_BACKEND_HERDR_BLOCKED_CONFIRM_DELAY).
2. Re-query the live agent status via herdr's agent.get API.
3. If the agent has already resumed working (or gone idle/done),
   absorb the transition silently -- this was a transient false alarm.
4. If the agent is still blocked, escalate as before -- this is a real
   wedge (permission dialog, stuck prompt, interactive menu).

This handles all three cases correctly:

• Transient prompts (1-2s): after 3s the agent is working -> absorbed.
• Long-running commands (sleep 600): herdr correctly reports working
  during command execution, so no blocked transition is emitted at all.
• Real wedges: the agent stays blocked for 3+ seconds -> escalated.

The 3-second delay is a deliberate trade-off: a real wedge is delayed by
3 seconds, but the firehose of false alarms (dozens per minute) is eliminated.
The captain sees a notification only when something is genuinely wrong.

Config

FM_BACKEND_HERDR_BLOCKED_CONFIRM_DELAY (default 3, in seconds). Set to 0
to skip the re-check entirely (used by tests that lack a live herdr server).

Additional change

Also adds a git-identity rule (rule 8) to the crewmate brief scaffold
(bin/fm-brief.sh). Workers sometimes invent an agent identity
(firstmate <firstmate@local>) when committing, instead of using the
captain's global git config. The new rule explicitly forbids overriding any
git identity field.

Testing

• apply_transition tests pass with FM_BACKEND_HERDR_BLOCKED_CONFIRM_DELAY=0
• Shellcheck clean on the modified files
• When the re-check API call fails (no herdr server, test environment), the
  function falls back safely to the old escalation behavior (fail-safe)