docs: the container never slept - the ~8.5min tail was a backfill artifact (#120)

[wormeyman]

The number in #120 was an artifact of reading the dataset too early

#120 published an ~8.5 minute idle tail, measured from containersUsageAdaptiveGroups about 14 minutes after the render that produced it. 22:50Z was recorded as absent - asleep.

Re-read now that the dataset has settled, that bucket is there, and so is every bucket after it. The placement that render woke - a20877a2, first bucket 2026-08-03T22:35:00Z - was still allocated 29.3 hours later:

five-minute buckets present	352 of 352
gaps > 5 min	none
distinct placements	1 (a20877a2)
total allocated memory	103,575 GiB-seconds
awake fraction (1 GiB instance)	0.981
worker requests in that window	5

Five requests cannot hold a container open for 29 hours against sleepAfter = "20s". There is no tail. The container never sleeps at all - which is precisely what the SIGTERM diagnosis in #154 predicts: @cloudflare/containers sends SIGTERM and never escalates to SIGKILL, node is PID 1 and ignored it, so the instance only ever goes away when a deploy replaces the placement.

Two cost claims in CLAUDE.md were wrong

1. The drain fix did not end the bill. Daily allocated memory, applicationId = a039bbb1, against the 96.0 GiB-hours a 4 GiB instance bills for a full day:

2026-07-13  95.6    2026-07-25  95.6    2026-07-31  95.7
2026-07-14  96.2    2026-07-26  97.0    2026-08-01  96.3
2026-07-15  98.2    2026-07-27  95.3    2026-08-02  99.0
2026-07-20  96.3    2026-07-28  96.1    2026-08-03  89.1 (downsize at 21:22Z)
2026-07-21  96.0    2026-07-29  87.0    2026-08-04  23.6 (= 24.0 at 1 GiB)
2026-07-22  96.8

100% of every full day, including the five days after the drain fix deployed on 2026-07-29. The ~$28/month was still being paid right up to the 2026-08-03 downsize, which cut it ~4x rather than ending it. The drain guard stays - the hazard is real and worker.spec.ts should keep testing it - but it was never the load-bearing cause of the bill.

2. The disk-to-memory ratios are dashboard units. Queried in bytes, which is what the GraphQL dataset actually returns, standard-1 reads 1.86 and basic reads 3.73, not 2.0 and 4.0. Both forms are now stated so neither reading surprises anyone.

What this PR does

Docs only - no behaviour change. It corrects those two claims and adds two bullets that were missing: the SIGTERM/PID-1 cause, and the backfill trap ("wait at least an hour before reading absence as sleep, and confirm with placementId continuity rather than bucket presence alone").

Status of the fix itself

Still not deployed. 343df53 is on main; production runs worker version 2e67470f and container app version 3 (image sha256:16205eda..., updated_at 2026-08-03T21:24:01Z), both of which predate the handler. #120 stays open until a preview:deploy and a settled re-read of the buckets.

Verification

pnpm run verify - exit 0. 1588 app tests passed / 3 skipped (196 files), 13 worker tests, 4 container tests including exits promptly on SIGTERM, 0 warnings.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QZvNS2H4cbaZk46ybA7hbj