fix(run): always run on schedule; rate-limit only the empty heartbeat

[krayong]

Problem

On a fixed-clock-time schedule (e.g. launchd StartCalendarInterval at 09:00), the digest fired only about every other day.

run_once gated the whole run on _catch_up_due — a strict now - last_successful_run >= INTERVAL_DELTA. Each run completes a few seconds after the scheduled instant, so the next same-time fire is always a hair under 24h and the gate skipped it entirely (no detection, no digest, run not marked). The skip left an empty 0-byte log and last_successful_run untouched, which is exactly how this was diagnosed on a live install.

Fix

• Remove the catch-up gate. Detection always runs when the native scheduler fires. Real changes were never the issue — the delivery ledger already makes them idempotent, so they always go out.
• Rate-limit only the empty heartbeat via _empty_digest_due: the "nothing notable" digest is suppressed when the previous successful run was under half an interval ago, so two fires close together (a manual run plus the scheduled one, or a wake double-fire) don't double-send it. --force overrides; cron always sends.
• mark_successful_run now fires on every run, so the window advances even when the heartbeat is suppressed.

Tests

TDD — the two new behavior tests were watched failing against the gated code first. Replaced the two obsolete catch-up tests with:

• test_run_within_interval_still_delivers_changes (the every-other-day fix)
• test_empty_digest_suppressed_when_last_run_recent
• test_empty_digest_sent_after_interval_with_prior_run
• test_force_sends_empty_despite_recent_run

Full suite: 472 passed; ruff clean. CLAUDE.md invariant updated.

🦇 Generated by Penny-One at Batcave