feat(evolution): cached-digest provider pre-flight for cron stages, model-resolution fix (#486)

[Lexus2016]

What

Restores the evolution cached-digest provider pre-flight as a tracked, tested feature and fixes the root bug that made it fail on production with no model configured for pre-flight ping.

• cron/evolution_preflight.py (new): a lightweight, non-streaming provider ping (max_tokens=1) plus a most-recent on-disk digest fallback for evolution pipeline cron stages (introspection / analysis / implementation / research / funnel / integration). Pre-flight is enabled by default and configurable via cron.preflight_enabled / cron.preflight_timeout_seconds.
• cron/scheduler.py: integrates the pre-flight in _run_job_impl, between runtime resolution and AIAgent construction. On ping failure it returns the latest stale digest as a graceful-degradation success, or raises if no digest exists. Purely additive — the sync: merge upstream/main (286 commits) — 2026-06-23 #487 native model-resolution path (job.model > HERMES_MODEL > config.yaml model.default) is untouched, and non-evolution jobs (stage is None) skip the block entirely.
• tests/cron/test_evolution_preflight.py (new): 27 tests covering stage detection, config, digest fallback, the ping per api_mode, and scheduler-level integration.

Why

When the configured provider is unreachable (e.g. Kimi timeout storms), evolution sessions burn retries/timeouts and deliver nothing. The pre-flight detects this quickly and falls back to the last good digest, so the pipeline keeps moving with stale-but-structured input instead of failing silently. This is graceful degradation, not a retry mechanism.

ROOT-FIX (this is why it failed on prod)

preflight_provider(runtime) reads runtime.get("model"), but resolve_runtime_provider() never populates a model key — the scheduler resolves the model into a separate local model variable and passes it to AIAgent(model=...) directly. On prod, runtime["model"] was therefore always empty, so the ping short-circuited with no model configured for pre-flight ping and the cached-digest fallback could never trigger.

Fix: before the ping, hand preflight_provider a runtime that carries the already-resolved model:

preflight_runtime = runtime if runtime.get("model") else {**runtime, "model": model}
err = evolution_preflight.preflight_provider(preflight_runtime, cfg=_cfg)

A shallow copy (not in-place mutation) keeps the ping side-effect-free, and the guard never clobbers a model the runtime may already carry (e.g. ACP-resolved). The local model is guaranteed non-empty at this point — _run_job_impl raises earlier if no model resolves.

Provenance

This feature previously existed only as untracked local code on the prod host; a git stash -u during a deploy hid it, and upstream-merge #487 overwrote scheduler.py with the native version without it. This PR restores it into git from that stash, with the root bug fixed.

Tested

• tests/cron/test_evolution_preflight.py — 27 passed. Includes a new unit test (test_resolved_model_does_not_bail_no_model) and a scheduler-level test (test_root_fix_runtime_model_synced_from_config_default) that reproduces the exact prod condition: resolve_runtime_provider() returns a runtime with no model, the job pins no model, and config.yaml supplies model.default. It asserts the runtime handed to the ping carries the resolved default.
• Negative control verified: that root-fix test fails without the fix (captured['model'] == None) and passes with it — so it genuinely exercises the bug.
• Adjacent suites green: test_scheduler_provider.py (29), test_scheduler.py model/runtime/run_job slice (77), test_run_one_job.py + test_codex_execution_paths.py (8).
• Syntax: ast.parse clean on all three files.

Closes #486.

[github-actions]

🔎 Lint report: evolution/issue-486-cached-digest-preflight vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 11397 on HEAD, 11396 on base (🆕 +1)

🆕 New issues (4):

Rule	Count
unresolved-import	3
invalid-assignment	1

First entries

cron/evolution_preflight.py:231: [unresolved-import] unresolved-import: Cannot resolve imported module `anthropic`
tests/cron/test_evolution_preflight.py:7: [unresolved-import] unresolved-import: Cannot resolve imported module `pytest`
tests/run_agent/test_credits_notices_toggle.py:76: [invalid-assignment] invalid-assignment: Object of type `None` is not assignable to attribute `_credits_session_start_micros` of type `int`
cron/evolution_preflight.py:204: [unresolved-import] unresolved-import: Cannot resolve imported module `openai`

✅ Fixed issues (2):

Rule	Count
unresolved-attribute	2

First entries

run_agent.py:3221: [unresolved-attribute] unresolved-attribute: Object of type `Self@get_credits_spent_micros` has no attribute `_credits_session_start_micros`
tests/run_agent/test_credits_notices_toggle.py:76: [unresolved-attribute] unresolved-attribute: Unresolved attribute `_credits_session_start_micros` on type `AIAgent`

Unchanged: 6028 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.