feat(FM.2): warm observe-only ambient monitor daemon - #699

Thanks @Maqbool61! The observe-only boundary and no-learning tests are useful safeguards. Before merging, please keep failed alert writes retryable, retain collector instances across ticks, remove exception payloads from logs, make single-instance admission atomic, and distinguish ambient alerts in the dashboard’s badges and counts. I reproduced lost alerts after a failed insert and duplicate admission during simultaneous starts. Once this lands, I’d love you to take #238 (level-5), which adds the collectors directly on top of your daemon.

Thanks for the detailed review, @fu351! Addressed all five:

• Retryable writes: record_decision now returns whether the row
  actually landed. The daemon's cursor only advances past a fully-durable
  batch — if any write fails, the cursor holds and the whole batch retries
  next tick rather than losing the alert.
• Retained collectors: run_forever discovers collectors once and
  passes the same instances into every tick instead of rebuilding them —
  collector-side state now survives across ticks.
• No exception payloads in logs: replaced exc_info=True with the
  error class name only (type(exc).__name__), matching the convention
  already used elsewhere in the codebase.
• Atomic single-instance admission: replaced the heartbeat-freshness
  check with an O_CREAT|O_EXCL lock file (same primitive
  storage/fingerprint.py already uses) — a single OS syscall, so two
  simultaneous starts can't both win. Verified with 8 real concurrent
  threads racing the same repo_root: exactly 1 wins, every time. A stale
  lock from a crashed process is still reclaimed via the heartbeat signal.
• Dashboard badges/counts: doberman dash's stats card and per-row feed
  badge now distinguish an ambient alert from a real BLOCK/AUTH (a separate
  "ALERT" badge/count), so the recent-BLOCK focal number isn't inflated by
  observations that were never enforced.

Repro'd your two bugs directly before fixing (lost alert after a forced
write failure, and the duplicate-admission race with concurrent threads) —
both are now covered by dedicated tests.

Merged, thanks @Maqbool61! Having the daemon check whether the row actually landed before it advances the bus cursor was the decision that made this safe, an alert can be retried but never silently lost, and the import-boundary test keeps the monitor from ever reaching the engine. Nothing changed in your code, I only used GitHub's update-branch to bring it level with main before the rebase merge. This closes #237. One gap I left for a follow-up, #703: doberman tune still reads ambient rows into the friction report. The next piece is #238 (level-5), the FM.3 collectors that feed the daemon you just built, would love for you to take it.