fix(compression): recover abandoned locks and suppress no-op retries

[the3asic]

What changed

• Reclaim a compression lease immediately when its structured pid= holder is
  proven dead on POSIX, instead of waiting for the full TTL.
• Keep Windows on TTL-only recovery because Python implements
  os.kill(pid, 0) destructively there.
• Stop the pre-API pressure loop from calling _compress_context three times
  when compression returns an unchanged message count.
• Add regression coverage for dead/live/legacy holders, Windows behavior, and
  no-op compression retry suppression.

Root cause

When a gateway process was killed during compression, its SQLite lease remained
valid for 300 seconds. The replacement gateway repeatedly saw the abandoned
lock and retried compaction. A second issue in the pre-API loop ignored the
documented no-op return contract and emitted three identical compact attempts in
the same turn.

Validation

• 88 focused compression/compaction tests passed on current upstream main
• ruff passed for all changed files
• Fable Max review: no blockers; PR-ready

[tonydwb]

Code Review Summary

Verdict: Approved (read-only token — formal approval requires write access)

Fixes compression recovery for abandoned locks and suppresses no-op retries.

• Addresses a race condition where compression locks could be abandoned
• Suppresses retry attempts that would be no-ops (nothing to compress)
• Well-scoped changes with clear documentation
• No security concerns

---

Reviewed by Hermes Agent

[teknium1]

The dead-PID lease reclamation half of this draft landed via salvage PR #70247 with your commits cherry-picked and authorship preserved — thanks @the3asic! Hardening on top: psutil.pid_exists as the probe (per CONTRIBUTING's os.kill(pid,0) ban; it's already a hard dep) and a same-pid self-reclaim guard. The attempts=3 loop hunk was dropped as discussed — it would clamp the user-configurable compression.max_attempts that landed via #69315. Closing the draft as landed.