docs(adrs): sweep the ADR corpus for inconsistencies and stale claims

[laurigates]

What

A 7-dimension sweep of the 9-ADR corpus and its dependent docs (roadmap, README,
hubs/, docs/api/, the blueprint manifest), adversarially verified. 50
candidate defects; 28 survived refutation and are fixed here.

The three classes it found

1. True when written, false now. The dominant class, and the one nothing
flags on its own.

Where	Was	Is
ADR-0008 Decision 3	"no training-step adapter, no VRAM read-out, no just bench"	PR #161 landed all three the day after the ADR
ADR-0004 M15 amendment	quant-path merge is "the tracked #83 follow-up"	#83 closed — tracked by nothing
ADR-0003 Decision 5	factory signature; eviction listed as an un-taken cut	both changed since

ADR-0008's paragraph is annotated, not overwritten — Decision 3 still binds,
because #110 is open and no real-model step has been timed. That distinction is
the whole point of the annotation.

2. Retracted in one place, still live in another. ADR-0005 Investigation
fact 5 asserted a ~1.5–2 GB shortfall that its own Addendum 2 measures at
~2.5–3× the card; Decision 3 still bolded the "fewer trained sites" lever that
its own preceding parenthetical retires. Both marked at the point of use, per the
corpus convention of annotating rather than deleting.

3. Unit-basis drift. The 10.3 GB vs 10.1 GiB MiB÷1000 artifact, reconciled
across ADR-0005, the roadmap and the example configs. The derived ≈ 24.7 GB sum
is deliberately left in the basis it was computed in rather than silently
re-footed — 10.1 GiB is 10.85 GB, so the arithmetic does not survive an in-place
number swap.

Also: 11 drifted code references repointed in the LoKr PRD (each verified against
the symbol, not the offset); ADR-0005/6/7/8/10 and PRD-0001 registered in the
blueprint manifest, which tracked only ADR-0001–0004.

What is deliberately not changed

ADR-0007 stays Proposed. The sweep flipped it to Accepted because ADR-0008
and ADR-0010 build on its Decision 3. A later ADR depending on one is not an
acceptance event, so it is reverted. The tension is real and worth a deliberate
call — see the gap list below.

ADR-0009 is not missing. It is reserved for #162's bench timing-mechanism
decision. No change needed.

Gaps this surfaces (need issues — not filed here)

1. Quant-path training-adapter merge is untracked. feat(trainer): optional Krea-2-Turbo training adapter (assistant LoRA merge-at-load) #83 carried it and is
   closed; the code still refuses training_adapter + compute.quant, which the
   int4 real-Turbo run needs.
2. The real Turbo training run is untracked. M15 shipped the load seam only;
   M14: end-to-end Krea 2 LoRA training + ComfyUI/Turbo interop proof #25 closed on Krea-2-Raw.
3. Evidence provenance. The 19.4 GB figure that ADR-0007/0008/0010, the README
   and the roadmap all quote has no committed artifact — the probe summary was
   never captured, the ledger logs are an off-repo popos path, and one
   measurement's instrumentation was on a deleted fork branch. The project already
   does this right once (docs/evidence/m14-krea2-dog-interop.jpg); the convention
   was never applied to the VRAM numbers. This is the same shape as the failure
   that forced ADR-0005's withdrawal sweep.

Verification

• just fmt-check, just lint, just test — green (66 suites, 0 failures).
• Link/anchor/manifest extractor: 0 broken links, 0 untracked ADRs.
• CommonMark render diff (markdown-it-py + pandoc), head vs worktree: structurally
  identical except the two intended additions.
• Every added issue/PR reference checked against its real state.

Three independent review lenses (render / facts / record) plus an adjudicator ran
over the final diff. They caught four defects the sweep itself introduced and
three in the hand-repairs, including a README claim whose stated discriminator
was wrong (burn#5162 needs the MMDiT graph — it is not about loaded vs random
weights, and on Metal the random-init case breaks) and two amendment markers dated
by event rather than by authoring. All corrected.

🤖 Generated with Claude Code

https://claude.ai/code/session_017syqaWwKyzpcgV5GB5ZrPw