[ShriD5]

Launch-time part of the idle-zero work: less work in the first minutes after launch, and no repeat of it on every relaunch.

Changes

• Salvage scan (5 s after launch) stops decoding every journaled chat.
  • Emptiness is the messages list's length (SessionDoc::has_message_entries), not a materialization of every entry.
  • A journal whose size and mtime match the boot that last found its doc healthy is skipped without loading the doc at all. The ledger is transcript-salvage.json, beside the journals dir. A doc can only lose its transcript in a session that also wrote its journal, so an unchanged journal needs no second look.
  • The recovery retry loop backs off between passes and ends once nothing is pending, instead of waking every 30 s for the life of the engine.
• Harness update checks wait and persist. The checks (--version, claude doctor, npm/GitHub, brew) start 90 s after launch instead of immediately. A complete check is cached in harness-update-status.json. A relaunch within the 6 h interval shows it, with dismissals re-applied, and schedules the next check for when it is due. Failed or partial checks are never restored, and an install clears the cache.
• Version probes run outside the global lock. This ports the version-probe part of upstream zeronsh/zeron 2e151ca6. binary_version held the global cache mutex across a CLI's --version, so one slow binary stalled every harness's resolution, and any failure (timeouts included) was cached for the binary's whole life. Probes now coalesce per binary. An inconclusive probe is retried after 30 s, a definite failure after 60 s, and an install racing a probe discards its answer. Our neutral probe working directory is kept.

Before / after

Same headless bench as #69: fresh data dir, 10 worktree chats, 2 PR watches, unattended client, 180 s settle, then 10 minutes measured.

	Before	After
Launch window (180 s) git / gh	151 / 6	151 / 6
10 idle min git / gh	375 / 14	375 / 14
Engine CPU (10 min)	2.85 s	2.94 s

The bench cannot show this PR. Its chats are fresh, with no journals, so the salvage scan has nothing to decode. Its PATH exposes no agent CLIs, so there is nothing for the harness checks to probe. The behaviour is pinned by tests instead:

• harness_updates::launch_waits_and_relaunch_reuses_a_fresh_check: no CLI probe at launch; a relaunch shows the cached answer with zero new probes; a dismissal survives the relaunch.
• harness_updates::stale_or_unsettled_caches_are_not_restored
• doc_host::salvage_sweep_remembers_healthy_journals, salvage_inspection_does_not_open_healthy_or_unrecoverable_chats
• schema::has_message_entries_reads_the_length_only
• Ported from upstream: executable::concurrent_probes_coalesce_per_binary_without_a_global_wait, timed_out_probe_is_retried_after_its_ttl_but_answers_are_kept, a_failed_probe_is_retried_once_its_failure_ages_out, a_probe_racing_an_invalidation_does_not_cache_its_answer, version_probe_bounds_hangs_and_rejects_nonzero

Local runs: zeron-harness executable tests 21 passed, zeron-doc 1 passed, zeron-engine harness_updates:: + salvage 39 passed.