What does this PR do?

This PR replaces a permanent-fingerprint latch in the kanban dispatcher's corrupt-board handler with time-bounded exponential backoff and a PRAGMA quick_check confirmation step. The latch is an in-process resilience heuristic (SECURITY.md §2.4); this change makes it less aggressive on transient I/O errors. Under the project's security policy, improvements to in-process heuristics are welcome as regular PRs (SECURITY.md §3.2).

Changes in scope

• gateway/run.py: Replace fingerprint-only latch in disabled_corrupt_boards with exponential backoff state dict (disabled_until_ts, backoff_seconds, fingerprint).
• Add _confirm_corruption(slug, exc) helper that runs PRAGMA quick_check before latching — if quick_check returns ok, the error was transient and no latch is applied.
• Add DEBUG-level log on every silent-skip tick (currently silent).
• Update error log on latch: include backoff duration + next-retry hint.
• Add success path: clear latch state on successful dispatch (disabled_corrupt_boards.pop).
• Add constants: INITIAL_BACKOFF_SEC = 30.0, MAX_BACKOFF_SEC = 900.0.
• New tests in tests/hermes_cli/test_kanban_dispatcher_resilience.py.
• AUTHOR_MAP entry in scripts/release.py.

Prior art & coordination

This change coordinates with PR #30410 (schema-drift fix in the same code block). If #30410 merges first, the implementer must rebase and handle the merge conflict carefully — keep #30410's disabled_schema_boards logic intact and apply the backoff change only to the disabled_corrupt_boards path.

Related Issue

Fixes #30417 (Bug 2: dispatcher resilience, transient SQLite I/O error causing permanent latch)

Refs #30410 (schema drift fix, same code block — coordinate on merge order)

Type of Change

• Bug fix (non-breaking change that fixes an issue)

Changes Made

• Modified gateway/run.py lines ~5118–5189 (_kanban_dispatcher_watcher / _tick_once_for_board):

  • Changed disabled_corrupt_boards dict type from dict[str, tuple] to dict[str, dict] with fields: disabled_until_ts, backoff_seconds, fingerprint
  • Added INITIAL_BACKOFF_SEC = 30.0 and MAX_BACKOFF_SEC = 900.0 constants
  • Added _confirm_corruption(slug, exc) function to run PRAGMA quick_check before latching
  • Updated _tick_once_for_board backoff logic to double backoff on repeated failures, reset on fingerprint change or successful dispatch
  • Added DEBUG log on silent-skip path
  • Added success path to clear latch state
  • Added import math if not present

• Added comprehensive test suite in tests/hermes_cli/test_kanban_dispatcher_resilience.py:

  • Test transient errors are not latched when quick_check passes
  • Test genuine corruption latches with initial backoff (30s)
  • Test backoff doubles on repeated failures, caps at 900s
  • Test fingerprint changes clear the latch immediately
  • Test successful dispatch clears latch state
  • Test backoff resets when fingerprint changes between failures
  • Test non-corruption errors bypass the latch mechanism
  • Test boards are not dispatched during active backoff window
  • Test malformed quick_check results are treated as corruption

• Added AUTHOR_MAP entry in scripts/release.py for steveonjava

How to Test

scripts/run_tests.sh tests/hermes_cli/test_kanban_dispatcher_resilience.py -v

All 9 new tests should pass. For full validation:

scripts/run_tests.sh
ruff check . && ruff format --check .
scripts/check-windows-footguns.py

Checklist

• Read CONTRIBUTING.md
• Conventional Commits
• Searched for existing PRs
• PR contains only related changes
• pytest tests/ -q passes
• Tests added for changes (required for bug fixes)
• Tested on platform
• Updated docs/config if needed