Does the phase apparatus need to shrink? (option B, deferred to run 2)

[mmcky]

Option B from #21 defect 4, deferred deliberately. Option A — correcting the duration claims and re-deriving doctrine §4's justification from evidence — is done in #24 (audit 0.1.3). This issue is the second half: does the phase apparatus itself need to shrink?

Why the question is open

The old justification for checkpointing was "bulk audits outlive sessions". Run 1 refuted it — 230 items in 22 minutes, none of the three named mechanisms in play. The rule survived on three replacement grounds, of which the strongest is that the checkpoint is evidence: run 1's 56-of-56 coverage reconciliation was done against findings.md rather than against the report, and the catalog was assembled from it.

That new justification is narrower than the one it replaced, and it does not cover everything the plugin currently does. If a checkpoint earns its place by being read, then a checkpoint nothing reads earns nothing.

The concrete case

Phase 3 wrote links.md (6 KB) at 16:58. 03-issue-links.md superseded it at 17:06 — eight minutes later, in the same session, and the bundle's own index says so: "Phase-3 working checkpoint, superseded by 03-issue-links.md." Nothing read it in between. On the evidence standard §4 now states, that artifact earned nothing.

Phase 2's findings.md is the opposite case and should clearly stay.

What to decide

Candidate	For	Against
Drop phase 3's named checkpoint	Never read; superseded in-session; §4's own criterion says it earns nothing	One data point. A run that dies between phases 3 and 4 loses the graph
Collapse phases 3 and 4	The relate step feeds the write step directly; the split buys a boundary nobody resumed from	Loses a clean division that made the procedure easy to describe
Drop the "Resumable from" table in §4	Describes a resume path never exercised	It is documentation of intent, cheap to keep
Keep five phases, shrink nothing	The division is advice, not contract (CATALOG § Principles); phases cost nothing if artifacts are optional	Leaves an artifact the doctrine's own criterion does not justify

Do not decide this yet

Wait for run 2. One measurement cannot distinguish "this apparatus is over-built" from "this repo was easy". QuantEcon/meta (161 open issues, no code to verify against) is the sharpest second data point in #16's matrix — and it is the run most likely to push the per-issue rate up, since org-wide issues resist the verify-against-the-branch method that made run 1 fast.

Two things would change the answer:

• If run 2 is slower per issue, the interruption argument strengthens and shrinking looks premature.
• If run 2 also never reads its phase-3 checkpoint, that is two independent runs agreeing, and the artifact should go.

Also relevant: run 2 should be the one that finally tests resumability, which run 1 left unvalidated. A resume that actually exercises findings.md is direct evidence about which checkpoints matter — and if resuming turns out to work from the snapshot plus phase 2's log alone, that settles phase 3 without further argument.

A note on who is arguing

The checkpoint machinery was introduced in #17 on the strength of §4's old premise. The same premise has now been retracted. That is a reason to hold this question open until there is more evidence than the argument that produced the machinery in the first place — not a reason to reverse it reflexively in the other direction.