docs(followup): retract an unsupported rate, restore two lost entries, keep caveats with their numbers

[JustMaier]

Docs-only. No code, no behaviour change. Three corrections, one root cause: a number outliving the evidence for it.

1. Retracts a rate that was quoted from n=2

The entry Verifier orphan detection is only ~50% sensitive had the retracted number in its title. The "~50% / ~1-in-2" barrier-timeout rate was n=2; as n grew it fell (1/2 → 2/5 → 2/6 → 2/7) toward ~25-30%, and no n worth quoting exists yet.

Rewritten with no rate at all. The argument never needed one — a barrier that misses even 1-in-4 publishes is a defect, and the exact rate changes nothing about the fix. Quoting one is how it spread; ava (who produced the figure) retracted it herself.

2. Removes advice the counter cannot support

The entry said to size the work from inconclusive_total's rate. It can't. That counter fires only on barrier_ok=false AND reread_present=false; reread_present=false has occurred zero times, with zero confirmed drops ever. inconclusive_total = 0 cannot distinguish "the barrier hid nothing" from "there was nothing to hide" — an ambiguous quiet counter, unresolvable by waiting, because the resolving event has never occurred. It is strictly dominated by a T+10s re-check (~20× the margin, no stall), so v1.1.49 deletes it.

Also records the finding that reshaped the v1.1.49 plan: deleting the barrier is a SWAP, not a subtraction. The ring is a bare VecDeque<u32> with no time dimension, so a naive "just requeue" yields ~0 delay — a hot re-check loop that reads pre-publish state exactly like the barrier and re-drives false drops. The scheduler is the work.

3. Moves a caveat next to the number it qualifies

The [flush-slow] promote= caveat (the fields don't sum: total=142ms with promote=184.9ms) lived one entry away from the 158-208ms figure it invalidates. So the freshness entry read as fact.

That split is not hypothetical: the figure got laundered into an N=10s design justification it cannot support — by me, hours after I wrote the caveat. A caveat that doesn't travel with its number isn't a caveat. The 93-98% timeout rate is unaffected: it's counted directly (13/14, 112/114) and never rested on the promote field.

Also: restores two entries lost in the v1.1.47 tag recovery

A dirty FOLLOWUP.md caused the v1.1.47 checkout abort → mis-based tag. Discarding it to re-cut cleanly dropped two entries, preserved and restored here:

• CI never compiles tests/ — 29 files / 178 #[test] fns never build or run; tests/bulk_load_fixture_test.rs is already rotted (imports a symbol pg_sync doesn't export) and CI has been green throughout. Coverage theatre is worse than no tests, because it's trusted.
• /debug/memory mixes stable and noisy fields — measured spread table; two engineers drew opposite false conclusions from single samples within ten minutes.

🤖 Generated with Claude Code