Close Stage 5: exit audit, eight overstated verdicts, and PyFlow 0.1.0

[AdamClemens]

A separate auditor-stance pass (prompts/common/AUDITOR.md) over the
thirteen-criterion verdict table TASK-034's own session wrote as it
closed. Eight of the thirteen did not survive, and one of the eight
was a defect in shipped behaviour rather than only in a verdict.

Per-criterion record: docs/planning/roadmap.md's new Status as of
2026-08-29 section, placed where all four prior stages put theirs —
the generated status report had been flagging its absence on its own
("has no pending tasks recorded, but isn't marked complete").

Seven failures share one shape

Every one was a later clause of a criterion whose first clause was
genuinely met — a sentence the criterion had already written down.

Criterion	Claimed	Actually
13	Met	Names two ADR-003 substitution checks; only the PressureCoupling one existed. register_linear_solver was never called outside assembly.py, so a PISO constructing its own solver would have passed every scenario in the repository. Confirmed by mutation, not argued.
10	Met	Six stale claims, in six files no Stage 5 task opened. README.md said "Stage 5 — not yet started" on the day Stage 5 closed. docs/architecture/sequences.md — whose only stated job is "in what order do things actually happen when PyFlow runs" — contained no mention of pressure, predictor, corrector or navier_stokes_step at all.
11	Met	Neither mvp.md nor releases.md recorded the MVP as reached, though releases.md names "Reaching the MVP" as a trigger it was told to act on the moment it fired.
6	Met	Enumerates five rejection surfaces; four existed. A periodic boundary carrying a prescribed velocity or per-field override loaded cleanly and was then ignored outright.
5	Met	"The absolute tolerance is stated and defended in the feature file against the mesh actually used" — none was stated. Errors of 10, 5 and 2 would have satisfied monotonic decrease exactly as well as the real ones.
7	Met	The Ghia cavity fixture is square, at origin (0,0), lid speed 1.0 — three violations of the degenerate-fixture rule, taken and unrecorded.
3	Met	The initial-divergence clause — the one that stops the two assertions after it being vacuous — was true (1.85 against 1e-4) and asserted nowhere.

The eighth was real

simulation.velocity_solved meant two different things. With no
scalar_pattern, _add_solved_velocity_rendering called
navier_stokes_step. With one, _add_passive_scalar_transport
transported velocity's components like any other scalar and never
pressure-corrected them — so a configuration saying "solved" produced a
velocity that was not incompressible, with no error and nothing rendered
differently, selected by whether a scalar happened to be configured.

Measured rather than argued — maximum divergence over 1, 10 and 40
frames:

	1	10	40
uncorrected	9.16	8.24	6.95
corrected	2.30	0.47	0.057

TASK-031 and TASK-034 both wrote the gap down — in two CLAUDE.md files,
and against no criterion, which is how it survived Criterion 12 being
marked met. docs/practices.md gains the rule that distinction produced.

Two checks that read like gates and were not

• README's Current Phase had gone a full stage stale for the second
  time (the Stage 2 audit found the first). Now mechanical:
  generate_status_report.py's drift check fails make ci when the
  stage README names is not the roadmap's own frontier. No second source
  of truth — the roadmap decides, README is checked against it.
• The Gherkin-scenario drift rule had been silently inert. Its
  pattern needed a literal space before "are Gherkin scenarios"; a later
  edit hard-wrapped the roadmap's line between the count and that phrase.
  A pattern that stops matching reports nothing to check, which reads
  exactly like a clean pass — the roadmap sat claiming 79 scenarios
  against a live 94, behind a green gate. All three claim patterns are
  now whitespace-tolerant, with a regression test quoting the real
  wrapped sentence.

Four maintainer decisions, all taken the thorough way

• PyFlow 0.1.0. releases.md rewritten with a real process —
  semantic versioning with MINOR carrying stage completion, a release
  cut only after a stage's exit audit completes, an annotated tag on a
  two-platform-green commit, nothing published anywhere yet (a recorded
  decision, not an omission). planning/data/releases.yaml populated for
  the first time; its own stated trigger had fired.
• Criterion 6's missing surface: built. A fourth rule in
  _validate_boundary_conditions_jointly, scoped to non-default values
  so every periodic configuration already in the repository stays valid.
  Six rejection tests, one acceptance test.
• Criterion 7: fix what can be fixed, record what cannot. Ghia's
  Re = 100 profiles are nondimensionalised on a unit square at unit lid
  speed — two forced exceptions, now stated in the feature file. The
  origin was never forced, and the cavity now runs at a non-trivial one.
  Verified by a full three-resolution run first: identical errors
  (0.1433, 0.0874, 0.0578), vortex centre 0.0010 from Ghia's own.
• velocity_solved: routed through. Both live paths call
  navier_stokes_step, with a regression test whose bound was chosen
  from both measurements above.

Verification

make ci green locally: 688 tests, 95 scenarios across 21 feature
files, every check passing (14m07s — the Ghia validation is ~13 of
them). Stage 5's own Criterion 9 is discharged by two real runs, both
green on ubuntu-latest and windows-latest: 33269489214 (PR #47) and
33270312866 (its merge to main).

This branch's own CI run is what the Merge Gate requires before it
merges — stated rather than assumed, since a local pass is not that
evidence. The v0.1.0 tag is deliberately not cut here: the process this
PR writes says a release is tagged on main after the stage's exit audit
completes, which is after this merges.

🤖 Generated with Claude Code