docs(soak): close Phase 6 24-h re-run — 9/9 assertions PASS end-to-end

[githubrobbi]

Summary

The last remaining v0.6.0 24-h-soak gate now closes against a live Windows-host capture.

LOG/uffs_soak/phase6-20260514-122946/ ran for 24 h on the 7-drive reference box (2026-05-14 12:29:46Z → 2026-05-15 12:29:46Z) against the post-PR-218 harness fix (`shard.ttl=debug` → `shard.ttl=trace`). Validator reported 9 of 9 assertions PASS end-to-end.

Contract evidence

Contract	Evidence	Status
C never demotes below Warm	0 `to=Parked` events for letter=C; 2 870 `Demote target clamped by per-drive min_tier` debug events	✅ end-to-end
Peer drives demote Warm → Parked	D / E / F / G / M / S each fired 2 transitions	✅ end-to-end
Adaptive TTL bonus engages	C.max_warm_ttl=3 786 s vs peer max=300 s (12.6× bonus)	✅ end-to-end (was §4.5b ⚠️)

Memory trajectory

```
00h 23h post-load
Working Set (WS) : 6 746 800 128 → 22 888 448 → 69 238 784 (308× WS trim, 3× post-load re-page)
Private Memory : 8 293 457 920 → 1 791 188 992 → 1 669 558 272 (78 % real release as drives demoted)
Virtual Memory : 28 172 120 064 → 28 168 974 336 → 28 168 974 336 (flat — no address-space leak)
NPM (non-paged) : 26 736 → 26 328 → 26 328 (flat)
```

The 78 % PM release is materially different from §4.5d ws-trace (where pm_bytes stayed within 3 % because the keep-warm worker held all 7 drives Warm). Here the controller actively demoted the 6 peer drives, which unloaded cold shards from memory — exactly the intended tiering behavior. The end-of-soak synthetic-load window on C re-paged recently-needed shards back into WS without re-allocating PM (PM actually continued to fall slightly).

No `panic` / `OutOfMemoryError` / `FATAL` log lines across the 24-h window.

Why the re-run was needed (§4.5b deferral root-cause)

The pre-fix soak (`phase6-20260509-213122/`) was filtered by `RUST_LOG=shard.ttl=debug`, which dropped every `below-ttl` TRACE event — the only event class carrying the bonused `warm_ttl_sec` field during the synthetic-load window (because drive C sits in Warm/Hot with `idle_secs ≈ 0`, so the demote-eval ladder never reaches either DEBUG arm).

PR #218 raised the harness filter to `shard.ttl=trace`. This re-run is the first 24-h soak captured against that fix.

Docs updated

File	Change
`docs/architecture/memory-tiering-windows-host-validation.md`	New §4.5e closing capture sub-section. §4.5b's closing paragraph forwards to §4.5e. §2 criterion 3 cross-ref picks up §4.5e.
`docs/architecture/memory-tiering-bake-criteria.md`	§3 Phase-6 checkbox ticked `[x]`; status text refreshed; cross-refs §4.5b + §4.5e.
`docs/architecture/memory-tiering-readiness-validation-2026-05-05.md`	Remaining-items list restructured: Phase 6 removed (closed); paper-trail subsection added for Phase 6 / Phase 7 / ws-trace closures; §6 closure note refreshed.
`CHANGELOG.md`	New "Verified — Phase 6 24-h Windows-host soak closes end-to-end (2026-05-14/15)" entry under Unreleased, with gate-status table showing all three v0.6.0 24-h-soak gates green.

v0.6.0 gate slate

Gate	Source	Result	Closed
Phase 6 (`min_tier=Warm` floor + adaptive bonus)	`phase6-20260514-122946/`	9 / 9 PASS	2026-05-15
Phase 7 (USN-journal churn)	`phase7-20260510-214412/`	7 / 7 PASS (regex fix)	2026-05-13
ws-trace (Working-Set trajectory)	`wstrace-20260513-113344/`	4 / 4 PASS	2026-05-13

Only the one-week `main` bake remains.

Discipline notes

• Pure-docs PR; no Rust changes.
• Atomic single-purpose commit, GPG-signed.
• Pre-push gates green locally (lint-fast: 3s; lint-pre-push: 4s).
• No suppression hacks; no tests touched; no public-API changes.