fix(weekday-sf): raise MorningEnrich timeout + poll-cap (L4552b) and force lib-pin re-resolve (L4591)

[cipher813]

Summary

Fixes two weekday-pipeline reliability defects in infrastructure/step_function_daily.json + scripts/ensure_lib_pin.sh. Scopes cipher813/alpha-engine-config#970 (L4552b) and cipher813/alpha-engine-config#1059 (blockers 2 + 3).

NOTE: merging this PR auto-deploys the weekday Step Function via deploy-infrastructure.yml (push-to-main) — the PR is the safe artifact; do not self-merge.

FIX 1 — L4552b MorningEnrich SSM timeout (#970, #1059 blocker 2)

The 6/11 weekday RunMorningEnrich timed out against the exact executionTimeout=1800 ceiling (SF state TimeoutSeconds=1860). Per the #970/#1059 diagnosis, daily_closes+intraday collection is legitimately >30 min on the t3.small, not hung — the slow ArcticDB append was already split to MorningArcticAppend (#983), so this is the remaining fetch tail.

• Raised SSM executionTimeout 1800→3000s, SSM-param TimeoutSeconds 1800→3000, SF state TimeoutSeconds 1860→3060 (60s margin).
• Added a bounded poll-iteration cap on the CheckMorningEnrichStatus loop (InitMorningEnrichPoll + IncrementMorningEnrichPoll + a ≥210-attempt branch, ~57 min — just past the 3000s ceiling) that fails fast into a new MorningEnrichPollTimeout state with a clear cause (routed through HandleFailure so the SNS failure alert carries it) instead of spinning to the state TimeoutSeconds on a stuck-InProgress SSM agent (the #970 symptom). Mirrors the existing InitSSMPollCounter/SSMReadyChoice/IncrementSSMPoll cap pattern already in the SF.

FIX 2 — L4591 weekday lib-pin auto-heal re-resolve gap (#1059 blocker 3)

The 6/10 weekday run crashed ImportError: cannot import name 'guard_entrypoint' (box ran new code against a stale installed lib).

Finding (reality vs the #1059 hypothesis): ensure_lib_pin.sh is already wired into the weekday MorningEnrich entrypoint between git-pull and the collector invocation — it landed in data#386 (the same 2026-06-10 fix) and is structurally guarded by test_sf_lib_pin_self_heal_wiring.py. So the wiring is present, not missing. The genuine remaining gap is the heal mechanism itself: a plain pip install of a git-URL pin treats the version as already-satisfied and skips the reinstall, so a new symbol (guard_entrypoint) or newly declared extras never re-resolve onto the long-lived box. Hardened the heal-path install to pip install --force-reinstall --no-cache-dir so the pinned ref's actual contents (symbols + extras) always land. (The Saturday data path is unaffected: it runs on ephemeral spot boxes that pip-install fresh each launch, which is why drift only bites the long-lived weekday EC2.)

Validation

• step_function_daily.json is valid JSON; no dangling/orphan states (graph checked).
• ensure_lib_pin.sh passes bash -n.
• Full tests/ suite green: 1981 passed, 2 skipped (pre-existing FutureWarnings only). SF wiring tests (weekday skipgate, lib-pin self-heal, morning-enrich split, poll resultselector) all pass.

References: L4552b, L4591, #970, #1059.

🤖 Generated with Claude Code