fix(e2e): probe prover via CPU health sentinel, not the GPU service

[alexgarden-mnemom]

Summary

• Repoints tests/e2e/config.js prover.base at the CPU-only health_sentinel URL (https://mnemom--mnemom-prover-health-sentinel.modal.run).
• Tightens the per-endpoint timeout from 60s → 10s now that we're off the GPU cold-start path.

Why

The deploy gate's prover health probe was hitting the H100 prover_service, which cold-starts in 60-180s (container boot + SP1 init, see modal_app.py:135-138). Any probe arriving on a cold container busted the 60s timeout and failed the gate, blocking every other repo's production deploy.

mnemom-prover PR #32 (Apr 28, 2026) added a CPU-only health_sentinel specifically for this purpose: it answers in milliseconds from a CPU container (scaledown_window=300s) by querying Postgres for proofs stuck in pending/proving > 10 min. That's the right readiness signal for the deploy gate — real outages flip it red, quiet idle periods leave it green.

Companion change

mnemom-prover PR #40 reverted the GPU prover_service scaledown_window from 5s back to 60s, restoring dispatch + GPU probe paths. This PR removes the deploy gate's dependence on that knob entirely by getting off the GPU probe path.

Verification

Probed the sentinel URL before opening this PR:

$ curl -sS -m 10 -o /dev/null -w "%{http_code} (%{time_total}s)\n" \
    https://mnemom--mnemom-prover-health-sentinel.modal.run/health
200 (9.3s)

9s on a cold sentinel; subsequent warm probes will be sub-second.

Test plan

• CI green on this PR
• Post-merge: trigger any service deploy (e.g., a no-op mnemom-api redeploy) and confirm the prover health check passes in <10s
• Re-run a previously-failed mnemom-api deploy from yesterday's incident and confirm it now succeeds
• Check Grafana: prover.dead_letter_breach rate stays at 0 (confirms PR Fix deploy.yml: add missing npm-directory for SDK publish jobs #40 fix is still holding)

Rollback

git revert this commit. Prover-side PR #40 ensures the GPU probe path also works, so rollback is safe — deploys would still pass, just paying GPU cold-start latency on quiet probes.

Follow-ups

• Verify BetterStack uptime monitor target is the sentinel URL (separate manual check; PR Fix PR deploy context token usage #32's design assumes it but unconfirmed today).
• Larger architectural follow-up: split this shared E2E suite into per-service smoke gates + a scheduled system-health cron, so one service's regression can't block another's deploy. Tracked in the prover repo's incident postmortem plan as PR-05.

ops-impact: high

Changes which endpoint every service's deploy gate probes. Soak: validated via direct curl pre-merge. Rollback path is one-line revert.

🤖 Generated with Claude Code