The 61-minute cold-mirror hydration floor is cited to #36, which does not contain it — and cloud#2681 says it was never measured

[khaliqgant]

Summary

Three source locations treat a 61-minute cold-mirror hydration as a measured
production fact and cite it to #36. It is the lower bound on the sweep
budget — the dial that decides whether a wedged discovery sweep can ever be
killed — so it is load-bearing on an active outage decision.

#36 does not contain that measurement, and the repo that owns the thing being
measured has an open issue stating it has never been measured.

Where the claim is made

• src/config/schema.ts, in the DEFAULT_DISCOVERY_SWEEP_BUDGET_MS block:

  "a budget below realistic cold-mirror hydration (fix(factory): replay existing Slack triage answers #36 measured 61 minutes in production)"

• src/orchestrator/factory.test.ts:17312, on a MUST-NOT-FIRE test:

  "fix(factory): replay existing Slack triage answers #36 measured a real cold-mirror reconcile at 3,665,173 ms (61 minutes) in production"

• src/mount/relayfile-operation-timeout.ts, repeating the constraint with no citation:

  "below realistic cold-mirror hydration a slow boot becomes a crash loop"

The DEFAULT_READINESS_RECONCILE_TIMEOUT_MS comment sizes the 90-minute default
against it directly: "90 minutes leaves roughly 47% headroom over that
measurement."

What #36 actually is

ref	actual title
AgentWorkforce/factory#36	MERGED PR — fix(factory): replay existing Slack triage answers
AgentWorkforce/factory-cloud#36	BLOCKER for enabling short sleep: backstop and event alarm can claim the same wake concurrently

Neither is a hydration measurement. Searching the org for the figure itself:

$ gh search code '3665173 org:AgentWorkforce'
  → zero results

The number 3_665_173 appears nowhere in the organisation except the single
test assertion at factory.test.ts:17317 that is justified by the comment
citing #36. The assertion and its own justification are the same source.

The owning repo says it was never measured

AgentWorkforce/cloud#2681 — open since 2026-08-15 —
"Instrument the relayfile-mount boot sync: it is on every delivery's critical
path and has never been measured." From its body:

"nobody has ever timed it. There is no elapsed log, no histogram, no p95 — in
cloud or relayfile. Every architectural decision about the mount (lazy vs
eager, poll vs FUSE, warm pool sizing, whether to keep the box at all)
currently rests on a number that does not exist."

Why this matters right now

The sweep budget is the fix for the current dispatch wedge (#372, #374, #376).
Choosing its value requires a defensible floor, because schema.ts documents
that the sweep commits its checkpoint only at the end — a timeout is
destructive, so a budget set below true hydration cost converts a slow boot into
a loop that never progresses.

If the real hydration cost is materially lower than 61 minutes, a much tighter
budget is safe and the outage has a cheap fix. If it is genuinely 61 minutes and
container lifetime is below that, then no budget value is both above the cliff
and low enough to fire, and the fix has to be a resumable checkpoint rather
than a deadline. These are opposite engineering conclusions and the evidence
currently cannot distinguish them.

This is not a claim the cliff is imaginary

The number may well have been measured in a run log or transcript that never
reached an issue. The defect is that it cannot be cited as measured in its
current form, and the citation is wrong wherever it points.

Asks

1. Correct or remove the #36 citation in all three locations — point it at the
   real evidence, or mark the figure explicitly as an unverified estimate.
2. If the measurement exists, attach it here so the constant has a source.
3. If it does not, treat AgentWorkforce/cloud#2681 as a blocking dependency for
   any evidence-driven tightening of sweepBudgetMs, and say so in schema.ts
   where the next person will read it.
4. Consider whether the factory.test.ts:17317 assertion should be relaxed or
   re-anchored, given its justification is currently circular.

[khaliqgant]

Second location found, same unsourced figure.

src/orchestrator/public-health.ts:54-58 restates the same 61-minute cold-mirror claim to justify READINESS_RECONCILE_STALL_INTERVALS = 10. So the number gates two constants, not one — the sweep deadline and the stall-reporting interval. If it is retired or re-cited, that second derivation needs redoing too.

Related correction from live measurement, since it bears on how this number gets used: the same schema.ts comment block reasons that each restart begins from zero because container disk is ephemeral. Across sweeps that is false — FACTORY_STATE_BACKEND: 'cloudflare-do' puts the discovery checkpoint in a Durable Object and #prepareDiscoverySessionreadsclaim.state.checkpoint, so it survives container restarts. What genuinely does not survive is the dependency PR probe walk, whose memo is cleared at the top of every #performRunOnce (factory.ts:3687`).

That matters here because the ephemeral-disk premise is part of what justifies sizing the deadline against a full cold hydration. If the checkpoint persists, the premise behind the 90-minute figure is weaker than the comment claims, independently of whether the 61 minutes was ever measured.

Measured mechanism is on #372 — the sweep is a livelock doing ~1,290 serial mount reads at 241 ms each, not a hang.