docs(0111): the live window is not starved — 0215 already fixed it

[karczuRF]

Summary

• Corrects this file's own claim that the live window "keeps up fine (2.81K unpriced)" — that figure is XLM-quoted only, conflated with the era table's all-quotes 10.44M of 17.05M (61.2%)
• Records a 48h partition-bounded re-measurement: the unpriced share is flat at 50-58% with no age-decay, so a bucket 44h old looks like one 4h old and enrichment never returns to those rows
• Splits the 26,900 rolling-4h total into the permanent exotic-quote floor (~45% in a long tail below rank 10, not this task's to fix) and ~2,370/4h of should-be-priceable USDC(3)/XLM(4) legs ≈ 14,200/day
• Documents why nothing alarmed: pass_metrics publishes only after pass.run() returns, the Lambda is killed mid-pass, and EnrichmentBacklogAlarm is treatMissingData: NOT_BREACHING — it reads OK on no data, the 0204 pattern
• Annotates AC 6, whose "floor" was the wrong number — judging against 2.81K would have read a working fix as a failure; the floor is ~27K falling toward ~12-15K, never zero

Docs-only. No code, no schema, no infra. The chosen design and phase ordering are unchanged — Phase 1's payoff is larger than recorded, because it un-starves the live window rather than merely speeding it up.

One thing deliberately left open and marked ⏳ in the file: whether legs 3 and 4 decay with age once the floor is stripped out. That query has not been run, so the starvation reading is stated as the likelier of two readings, not as established.

[karczuRF]

⛔ Correction — this PR's original conclusion is refuted by its own discriminator

The first version of this PR closed with "whether legs 3 and 4 decay with age is the one query not run — flat ⇒ starved, decaying ⇒ ordinary lag." That query has now been run. They decay.

Measured 2026-08-24, 48h in 4h buckets, partition-bounded, using enrichment's real CANDIDATE_PRED:

leg	candidates (40h, freshest 8h excluded)	unpriced	%
XLM (4)	138,722	12	0.0
USDC (3)	90,672	0	0.0
every other leg	~147,320	~147,320	exactly 100

There is no middle ground — 100.0% on every exotic leg, not 99% or 95%. That is a structural gap (no oracle, no pivot reference), not a backlog. It matches 0056's exotic-quote floor: ~88K rows/day against ~138K/day priceable.

Both measurements were correct — different regimes

The original 48h window ran 08-19 20:00 → 08-21 20:45. The 0215 Caddy fix landed before 14:00 on 08-21, so ~42 of those 48 hours are pre-fix, when the pass died at 30.0s on every invocation and the live window genuinely was starved.

The aggregates reconcile exactly: 08-22 20:00 gives 17,038 of 33,268 = 51.2%, squarely inside the original 50–58% band. The aggregate looks flat because the floor is half the candidates at ~99.7% and never moves, masking a priceable half that goes to zero. The original analysis identified that masking effect and then drew the opposite conclusion from it without testing.

The starvation was real and 0215 already fixed it. It was never 0111's scan cost.

What changed in the file

• The 08-21 section is kept, marked superseded-as-verdict/valid-as-record — it is the pre-fix baseline.
• "Phase 1's payoff is larger" is withdrawn (preserved with a banner; the reasoning is sound given a starved window).
• AC 6 rewritten leg-scoped. EnrichmentRowsRemainingRecent is ~100% exotic floor and reads the same whether the fix works or not — it would have passed a starved pass. The check is now USDC/XLM at ~0 outside the newest bucket.
• Leg identities confirmed against prices.assets (0139's warning heeded). USDT (111) is absent from the top 20 — negligible live volume.
• yXLM (44,432) and yUSDC (21,996) are 45% of the floor and belong to 0154. No task spawned.

Every reason Phase 1 is still needed is untouched: the alarm at 300,000/300,338 ms, 5.4 of 40 budgeted batches, the ~322-day drain, and 0218 blocked behind it.