M6d: axisymmetric relief, and reconciling the docs with the gates

[ggange]

Two commits: the M6d milestone, and a docs reconciliation that came out of reviewing it.

M6d — axisymmetric gas dynamics and radial relief (8c6732e)

M6c's G7 is ungated, and its spec justified that with one omission: "a planar 1-D code has no radial relief … it is the one effect the geometry has removed by assumption." This puts the effect back and measures it.

src/euler2d.rs is planar/axisymmetric Euler on annular cells, reusing euler1d's HLLC rather than carrying a second copy: a sweep packs (ρ, ρu_∥, E − ½ρv_t²) into the 1-D state and the transverse flux comes back as F_ρ·v_t with the upwind side read off sign(F_ρ). Both identities are exact. The area-weighted form means the 1/r never appears in the code — the axis interface has zero area, and the geometric source is written as the same floating-point expression as the pressure part of the flux difference, so a radially uniform state is a bit-exact fixed point. That is what lets G9 assert equality rather than a tolerance.

Verification: G9 planar limit reproduces Euler1d bit for bit over 240 cells × 40 steps; G10 Sedov–Taylor exponent 0.38628 vs the exact 2/5, level and peak compression gated as trends under refinement; G11 1.861/1.964 against a split-source contrast at 1.030/1.155; G12 conservation to <1e-13 with radial momentum deliberately not conserved plus an escape-flux leg; G13 axis entropy defect 2.99e-7 → 3.12e-8 under refinement against 3.02e-6 for an even-parity contrast; G14 wide-beam limit reproduces the 1-D column to 3.1e-13.

Sedov–Taylor is the repo's first multidimensional anchor, and ξ₀ is derived from the energy integral rather than quoted — 1.03278 against the published ≈1.033, so the literature value is a cross-check instead of an input.

The result (G15, pinned): radial relief costs δ = 1 − D/D_wide = 0.230 of the front speed at R_b·α = 3.2 and 0.305 at 1.6, monotone in beam radius. Pinned as a band of ±13 %, and that width is measured rather than chosen: grid +6 %, seed −7 %, ignition threshold ±8 %. Shown not to be a boundary effect (21.1/21.3/21.3 % at 3/5/8 beam radii). So relief is real and is not the whole ~2× gap to measurement — G7 stays ungated, but now for one reason only, the missing dataset.

The unlooked-for result: the modelled front is transversely unstable, growing cellular structure out of round-off, amplitude-proportionally, saturating at |u_r| ≈ 200–400 m/s, identically in planar and axisymmetric geometry. A 1-D solver structurally cannot show it, and it had to be separated from relief before the relief number meant anything — which is why δ is measured against the wide-beam 2-D run and not against the 1-D column.

Docs reconciliation (36f15de)

Reviewing the above turned up that MODELS.md has been written append-only: each investigation added a section and left the paragraphs above it describing a model that no longer ships. The claims ledger was kept current, so the drift hid underneath a table that looked right.

§ M6a prose said the seed was one electron in the focal volume and the multiphoton source off by default (both retired 2026-07-31 — the seed is q/ν_att and PPT is on), quoted n = 0.095 from the retired mean-trajectory closure, called the I_thr(p) slope gate RED five weeks after it went green, and carried Chylek local exponents of 1.951/1.047/0.170 against a current 0.501/0.857/0.386. Rewritten to the current model; the superseded narrative was already in M6A_SPEC.md under its own Superseded headings, so it is pointed at rather than retold (−89 lines).

The same drift had reached the source. breakdown0d.rs documented with_ppt_mpi as off by default (it is on) and with_seed_density's default as 1/V_focal (it is the ambient background); validation.rs's doc comment quoted an envelope of [0.183, 0.407] while its assertions twelve lines below pin [0.174, 0.382]. Comment changes only — no behaviour moved.

Audit. All 122 ledger rows were checked against the assert! in the gate each names. Three carried stale numbers; four that looked wrong turned out right, because several gates deliberately run seeding_suppressed to isolate one change. M1–M4, M6a.2, M6c and M6d rows all match their gates.

The census was wrong and nothing could catch it — the line claimed 118 rows (80/10/16/12) against an actual 122 (83/10/17/12). tests/docs.rs now parses the ledger, counts by status, and fails printing the corrected line; a second gate asserts every row carries one of the four statuses. Both were made to fail before being trusted.

M6c prose carried the same outward: G7 still justified by the excuse M6d retired; the lsd demo quoting a threshold of 1.14e16 that "does not fall with pulse length" when it is 8.815e15 at 6 ns converging to 6.745e15; M6a's ungated level as 4.8–7.0× in five places against a pinned 3.90–4.69×. The demo figures were re-checked by running the case: D = 5400.7 vs Raizer 5390.7 (+0.185 %), budget 1.26e-16, τ = 374.2 — all matched.

M6C_SPEC's copy of the threshold paragraph is amended in place with a dated note rather than rewritten, because the shape of the claim changed and the spec records what the milestone landed believing.

README's milestone table is compressed 30 KB → 17 KB; one cell held 10,305 characters. Every measured gate number and every ungated/pinned flag stays, per CONTRIBUTING.

Gates

All four pass locally, in order: cargo fmt --all -- --check, cargo clippy --all-targets --all-features under RUSTFLAGS="-D warnings", cargo build --all-targets, cargo test — 245 passed, 0 failed, 0 ignored.

Known limitation

The census gate catches structural drift only. A row whose prose quotes a stale slope is checked by nobody — the ground truth for any such number is the assert! in the gate it names, which CONTRIBUTING now says explicitly. Mechanising that would mean the ledger emitting numbers rather than quoting them.

🤖 Generated with Claude Code