fix(cron): restore cumulative runtime safeguards on current main

[jayleaton]

Summary

• restores the reviewed provider reset-window backoff and automatic cron resume from PR fix(cron): add spend-neutral provider reset backoff #102073
• restores task-bound Kanban worker guidance from PR fix(agent): bind kanban worker guidance to dispatched tasks #102857 so unbound cron sessions do not receive worker-only lifecycle instructions
• restores fire-fence and shutdown terminal durability from PR fix(cron): tri-state fire-claim heartbeat — fence contention is not ownership loss (c-027) #102627 so a completed, delivered run is not demoted by self-contention while genuine cancellation and takeover still fail closed
• preserves cumulative source head 7d030bace2240baaacd6ae603f2ccc57fcfa15e6 as the merge commit's second parent while retaining current-main content for separately owned lifecycle/dispatcher work

Provenance

• current-main parent: 79445a496c86a19332ad786494b8384d2167e2d0
• reviewed cumulative parent: 7d030bace2240baaacd6ae603f2ccc57fcfa15e6
• provider backoff: f40e3977a327b8c5c3f3edad7b957d75a0cf6a19
• terminal durability lineage: 9b26dcb9fb1f5425b198495b655d7ef54304da3f, 51e201b67397fb596cab6af88d5ac799f8d5122e, 991095136967b0f2776d451bde02dcc05b5ad0ea, d74a8d157a5a4e3fabac716df2d9c14420ec34ae
• task-bound protocol gate: 97ac4a6c5eb661cc963521cd9180db518cb28e01

Verification

• RED on current main: provider backoff 16 failures; Kanban guidance 3 failures; fire-fence durability 4 failures plus shutdown-state setup errors
• focused restored matrix: 318 passed, 0 failed
• broad cron/protocol/integration matrix: 1421 passed, 0 failed, 2 Windows-only skipped
• ruff, compileall, compat-pointer checker, and git diff --check pass
• broad matrix exercised schedule parsing, execution/delivery SQLite DDL, restart-safe external worker paths, scheduled worker claim/fence boundaries, and delivery terminal accounting

Boundaries

• no model/provider/spend/cap/security configuration changes
• no merge, gateway start/restart/reload, cron resumption, or product sign-off
• separately owned agent/kanban_stop.py, tools/delegate_tool*.py, and hermes_cli/kanban_db_dispatch.py are untouched
• deployment remains gated on independent review and explicit deploy authorization

[Enough1122]

AI code review — automated review for reference; please use your judgment.

Summary: Restores cumulative runtime safeguards on the cron path: task-bound Kanban guidance (fail-closed), clean-worker-exit lifecycle verification with durable read-back, fire-fence contention handling, and a durable provider rate-limit backoff policy. Large PR (~24 files) dominated by new tests (~1.3k lines).

Findings (Non-blocking):

• agent/system_prompt.py:31-46 — resolve_kanban_worker_guidance requires both the tool and HERMES_KANBAN_TASK and dispatcher-owned context, failing closed on ambiguity. Lifecycle instructions that mutate board state should indeed be task-bound, not merely tool-bound. The agent_init.py + _tool_guidance_block fallback both route through it, so lightweight agents bypassing _load_tools get the same gate. Good.
• cli.py:77-120 — _ensure_kanban_worker_lifecycle finalizes via one connection and re-reads through a fresh connection before reporting success; missing handoffs become durable protocol_violations instead of silent nominal exits. Retry delays (0/50/200ms) are short but this is a local-DB handshake, not network I/O. Fine.
• cron/jobs.py:204-222 — _acquire_flock now re-raises non-contention OSErrors instead of treating every lock error as "busy, retry". Previously a permissions/path error would spin until timeout; now only EACCES/EAGAIN/EWOULDBLOCK retry. Correct and strictly better.
• cron/rate_limit_backoff.py (new) — parses Retry-After / epoch-reset / human date formats with jitter and a 31-day declared-delay cap. Regex-based error-text parsing is inherently heuristic (a message mentioning "429" in passing could trigger backoff), but the consequence is only a delayed retry, and fallback bounds keep it sane. Non-blocking observation.
• Scope note: I reviewed the agent_init/system_prompt/cli hunks and sampled cron/jobs.py + rate_limit_backoff.py; the scheduler/provider/web-router/test bulk was surveyed by file list only. No blocking issues found in the reviewed portion; the remainder follows the same fail-closed pattern per the test names.

Verdict: Looks good. Nothing blocking.

[jayleaton]

Cumulative P0 provider-health continuation is now pushed on this existing PR branch at exact head de34ea656013d4365e38be40aa2b78eef89355c0 (implementation commit 9f18f10549b77039c162408b07fafbabb677ac49, parent f50bf4c80f28287b474058fb64b3689695f9946b). It adds one profile-scoped SQLite circuit shared by agent fallback/error/success, cron preflight/outcome, Kanban pre-claim dispatch, and staged reset-aware coder watchdog integration. Verification: 46/46 focused provider/fallback tests; 1406 passed, 3 platform-only skipped across full cron plus dispatch/tool integrations; Ruff, compileall, compat-pointer check, and diff check pass. Deployment remains unauthorized. Manifest: ops/t_5cf9660f-deployment-manifest.json. Hotspot: dispatcher overlap with active task t_ab2d132d must be reconciled before deployment.