Review follow-ups for [16]: persist the regulation stack, withdraw the unsourced 2005 load editions, distribute wind by height

[diegokingston]

Follow-up to the code review on #79. Three fixes, each with the regression test that failed first. Based on pr/16-cirsoc-2025-load-codes at 44d7c9a5, and targeted at that branch so it lands inside #79 rather than racing it.

Nothing here disputes the design of #79 — all three are cases where the code does not yet do what the branch's own comments and tests say it does.

1. The regulation stack never survived a save (5393dba1)

regulations and revisions are declared on StructureModel and on ModelSnapshot, but snapshot() built an explicit object literal that omitted both, and restore() never read them back. Every persistence path goes through that pair — .ded save/open, autosave, URL share, tab capture, and undo/redo — so the whole role stack and the staleness stamps were dropped on each one and the project silently fell back to defaultRegulations(). clear() did not reset them either, so a new project inherited the previous one's bindings.

The restore path now goes through migrateRegulations instead of assigning raw. That function was written for exactly this caller and never had one: a stored project naming a withdrawn edition comes back unset rather than bound to rules the app cannot apply, and a project saved before the stack existed migrates from the v1 codeSettings shape.

Worth knowing while reviewing the tests: the first draft of them passed against the unfixed code, because clear() leaves these fields in place and the stale value satisfied the assertion. They only turn red once the live fields are blanked between snapshot and restore, which is what blankLiveStack() is for.

This commit also adds imports for StoredRegulations / RevisionVector, which model.svelte.ts referenced without importing.

2. CIRSOC 101-2005 and 102-2005 were selectable but ran the 2025 rules (be34d9fd)

201-2005 is correctly reserved as UNAVAILABLE_SOURCE. 101-2005 and 102-2005 carry the same textAvailable: false in REGULATIONS, but had no availability field, so availabilityOf defaulted them to AVAILABLE, optionsForRole offered them, and roleUsable accepted them.

Nothing implements those editions. cirsoc102/wind.ts is a rebuild against the 2025 text with no edition parameter, and cirsoc101/combinations.ts and live-loads.ts hardcode 2025 in every clause ref. So picking "CIRSOC 102 (2005)" gave a project stamped 2005 whose wind loads were computed by, and cited, the 2025 rules — what optionsForRole's own docstring says the availability model exists to prevent. The note shown was only "Superseded edition, kept for legacy projects", which declares no substitution.

Two consequences handled here:

• the v1 migration bound those adapter ids directly and would now throw in bindRole; it binds the edition in force and emits editionWithdrawn per role, as it already did for concrete;
• roles-revisions.test.ts asserted the old contract in two places (one comment read "101-2005 and 102-2005 remain available"), so both are updated.

The guard test in edition-availability.test.ts could not have caught this: it asserted expect(o.role === 'concrete').toBe(false), which for a basis/loads/wind option reduces to expect(false).toBe(false) and passed vacuously for precisely the roles it was not covering. It now collects every offending option, asserts the list is empty, and names them on failure.

No UI change, same as when 201-2005 was withdrawn — the panel already lists non-available editions read-only with their reason.

3. Wind was not distributed by height (302d2e0c)

wind.ts evaluates the windward wall at [min(5, h), h] specifically "so the caller can distribute". The caller used .find(), which returns the first entry — the z = 5 m sample, the weakest point of the profile — computed the net pressure once outside the per-level loop, and applied it unchanged to every storey. The mean-roof-height value it had just computed was never read.

K_z = 2.41·(max(z,5)/z_g)^(2/α) rises monotonically with height, so this under-predicted wind on everything above 5 m: about 20% at z = 15 m and 30% at z = 30 m for 45 m/s in exposure C. The §1.10 minimum does not mask it for ordinary multi-storey buildings. It is unconservative and it reaches member sizing.

netAlongWindPressureAt lives in wind.ts beside the pressure rules rather than in the load generator, so the height rule keeps one implementation. Only the windward term varies with height — leeward is q_h over the full height per §2.4.1, and GC_pi cancels between the two faces.

Tests

3428 passed / 0 failed on a quiet machine, against 3435 on the unmodified branch plus the 6 new tests here.

One caveat on how that was measured: this needs a locally built WASM engine. Without npm run wasm the suite fails ~1100 tests on the branch as-is, which is an environment artifact and not a signal. Under concurrent load I also saw a handful of Test timed out failures appear and disappear on both the modified and unmodified branch — flaky, load-sensitive, and unrelated to these changes.

Not addressed

migrateRegulations returns notices, including editionWithdrawn, and nothing surfaces them. After this PR the bindings are right, but a user opening a project that named a withdrawn edition is still not told. Wiring that to a UI notification is a design decision that belongs to you rather than to a review follow-up, so I left it.

Separately, and not a code change: the ~230 new codes.* / regulations.* / loadPlan.* / maturity.* keys land in en/es only, so the other 12 locales silently render English via the dicts.en fallback. I understand that's the declared EN/ES scope. Flagging it because locale-parity.test.ts exists for this exact pattern and is scoped to design.*, so it does not cover the new namespaces — widening it would be cheap.

🤖 Generated with Claude Code