# V1-2.1 corrections after the independent audit: evidence Owner brief 2026-10-28 « V1-2.1 : corrections après audit indépendant » (verdict CHANGES REQUIRED on 3914eaa). Every run below used the isolated `pnpm db:sandbox` stack only (containers `faunosteo_dbsandbox_*`, ports 57xxx); DEV was not read or written. ## 1. Defects F1-F19 on the old schema (before M4) Script `2026-10-28-v1-2-1-repro-old-schema.sql` (one rolled-back transaction) on a sandbox of the 12 base migrations of develop, run during the first V1-2.1 pass. Output: ``` F1 plain account created a personal PUBLIC sheet for species other F5 an account with an UNCONFIRMED email created a sheet F6 set_sheet_visibility changes visibility with no gate F15 another plain account reads the public personal sheet: 1 row F9 editor rewrote draft content / renamed the sheet by direct UPDATE F4 version published with a drawing never confirmed F2 the draft of an ARCHIVED sheet was published; create_sheet_draft opened a draft on it F8 after archiving, another account reads 0 published versions F13 a DEPRECATED version was deleted physically F3 copied asset rows not confirmed F14 a platform admin at aal1 published an official sheet F19 grant_platform_role(owner) accepted at aal1; the new owner holds 0 verified factors F7 upload confirmed with no object in Storage F12 service_role published a draft by direct UPDATE F17 service_role inserted an owner row directly into profile_roles F16 sha256 of an asset used by a published version changed ``` Re-run on 2026-10-28 for the final decisions pass, raw psql output committed as `2026-10-28-v1-2-1-repro-old-schema.out` (sandbox `up fresh` from a copy holding only the 12 migrations before M4; the same NOTICE lines, `repro exit 0`, the transaction rolled back). F18 (last-owner race) was reproduced with two sessions; its test is `packages/supabase/tests/sandbox/sheets.sandbox.ts`. ## 2. service_role effective table privileges (§16.2 prerequisite) Measured on the M7 rehearsal stack (DEV path), before and after M4+M5: - before: ALL (DELETE, INSERT, REFERENCES, SELECT, TRIGGER, TRUNCATE, UPDATE) on `species`, `anatomical_sheets`, `anatomical_sheet_versions`, `anatomical_sheet_assets`, `profiles`, `profile_roles`; - after: none on those six tables (pgTAP 005 and 018 assert it, with the EXECUTE census). ## 3. New tests red on 3914eaa, green on the corrected HEAD Stack built from the base migrations plus the M4/M5 of 3914eaa, then the new and changed files of this pass: | File | On 3914eaa | On the corrected HEAD | |---|---|---| | 006 | 2 not ok: another editor uploads a pending key it did not register (L5) | 16/16 | | 016 | 1 not ok: a draft from a deprecated version takes it as base (M2) | 50/50 | | 017 | M1 approval on an archived sheet succeeds; M2 base from a deprecated version; the file aborts there | 85/85 (111/111 after the final decisions pass, §4) | | 018 | 93 not ok of 337: capability holders at aal1 get `rights.capability` instead of `rights.mfa` (§6.7); `delete_sheet_asset` answers a raw code | 337/337 | | 019 | 38/38 (lost coverage of behaviour that was already correct) | 38/38 | | sheets-concurrency | 5 failed of 15: publish against withdraw (no version lock), two deprecations (no sheet lock), archive against approve (no sheet lock) | 15/15 | ## 4. Real-stack results on the corrected HEAD Final decisions pass (owner brief « V1-2.1 : décisions finales du propriétaire »), code at `2379d28` (M6 option A, text size measure with the bounded, linear number walk, pgTAP 015/017/020; only `.claude/` differs at the final HEAD). Earlier figures (1 333 pgTAP) were the `2af61e7` state; intermediate passes at `7eca441` (1 352) and `5620399` (1 807) preceded the review findings below. | Run | Fresh stack | Upgraded stack (M3 + Git fixture, then M4+M5) | |---|---|---| | pgTAP (`test-db`, 18 files) | 1 808/1 808 | 1 808/1 808 | | sandbox suite, run 1 | 87/87 | 87/87 | | sandbox suite, run 2 (same stack) | 87/87 | 87/87 | | of which two-session concurrency (`sheets-concurrency`) | 16/16 each run | 16/16 each run | | pgTAP again after both suite runs | 1 808/1 808 | 1 808/1 808 | | `types --check` | types match | types match | PGlite: 128/128 (packages/supabase: the 18 pgTAP files and the closed-code census). New in this pass: - 017 M6: owner self-grants `sheet_reviewer`; owner-reviewer acknowledges and approves a version written by E, E publishes; owner refused on a round it authored (`review.separation`); publication after an author's approval refused (`publish.approval.invalid`). - 017 size: cap − 1 and cap saved and measured exactly; cap + 1 refused; a 2 MiB document of 3-byte characters saved; cap + 1 in 4-byte characters refused; a dense document within the cap as text and over it as binary jsonb saved whole; creation over the cap refused; numbers that JavaScript prints with an exponent stored and measured in plain form; a document under the cap in compact JavaScript form but over it as measured refused; 9 000 numbers of 131 072 digits refused with the closed code; 600 000 small numbers within the cap saved (the number walk is linear); the table CHECK definition. Every refusal answers `document.too_large` (23514), never a raw error. - 017 `publish.species.other` refused alone. - 020 (449): the AR-1 B matrix of §15 in full (6 caller classes × aal1/aal2 × grant/revoke × 4 roles × self/other, each refusal with « nothing written »), the regression chain on the real owner and admin at aal1, and the rights RPCs (`grant_catalog_right` on itself and on another, `revoke_catalog_right`, `lookup_catalog_account`, `list_catalog_grants`) refused to every non-owner class at aal2 (`rights.owner`) and at aal1 (`rights.mfa`). Number walk, linear (security gate on `7c4bf12`: `jsonb_path_query` row by row is quadratic on PostgreSQL 17). Raw psql timings on the sandbox, committed as `2026-10-28-v1-2-1-number-walk.out` (one transaction ending in `rollback`; script in its header line below): row by row 138 / 506 / 2 172 ms for 50k / 100k / 200k numbers; the shipped walk (one `jsonb_path_query_array` call) 69 / 135 / 267 / 421 ms for 100k / 200k / 400k / 600k. Over HTTP on both stacks, each run (lines extracted in `2026-10-28-v1-2-1-sandbox-timings.out`): 600 000 numbers (1 801 760 bytes of text) saved in 677-695 ms; 1 000 000 numbers answered `document.too_large` in 1 019-1 078 ms (budget 2 000 ms). Number bomb on the real image (sandbox, re-run on `2379d28`): the shipped function answers `document.too_large`; the same payload under the previous body (`octet_length(p_doc::text)` alone, rolled back) fails with `out of memory` after a 1 GiB buffer. ## 5. Timing at the caps (§15, budget 2 000 ms for save, submit, publish) Synthetic document: 64 views, 1 024 annotations, 8 192 finding nodes (depth 8), 1 089 review groups, 4 drawings of 2 000 000 bytes. The labels are calibrated on the contract's text measure alone: text 2 076 707 bytes (cap 2 097 152), binary jsonb 2 346 648 bytes, over the cap and accepted, as the measure requires. HTTP wall-clock through PostgREST on the development machine, code `2379d28`, milliseconds (min / median / max): | Stack, run | save (7) | submit (5) | approve (3) | publish (3) | |---|---|---|---|---| | fresh, 1 | 127 / 137 / 155 | 4 / 5 / 6 | 190 / 285 / 292 | 195 / 288 / 303 | | fresh, 2 | 130 / 142 / 155 | 4 / 7 / 13 | 196 / 284 / 297 | 205 / 302 / 306 | | upgraded, 1 | 135 / 139 / 148 | 5 / 6 / 6 | 192 / 282 / 294 | 196 / 292 / 296 | | upgraded, 2 | 127 / 137 / 151 | 5 / 5 / 7 | 199 / 286 / 295 | 220 / 292 / 299 | The core's at-caps fixture is V1-2.2 work; the contract fixes the budgets at the V1-2.3 checkpoint and measures verification (120 s) there. ## 6. M7: upgrade rehearsal with a Git fixture of the drawing object Accepted by the owner for V1-2.1 (brief « V1-2.1 : décisions finales du propriétaire », 2026-10-28; contract §12.3, §17.3). What this rehearsal is, and is not: - It uses a representative Git fixture, not a copy of the DEV object: the committed file `supabase/seed/anatomical-sheets/equine/draft/assets/horse.svg` (132 997 bytes), whose sha256 `0ea5202f…08541` equals the sha256 referenced by the DEV metadata (the asset row `7f767f74…` of the historical sheet, seeded from `equine_draft.sql`). - Its bytes, its Storage metadata and its relations (asset row, historical sheet, version) are compared before and after the migration. - No file was read from the real Storage of DEV, and nothing on DEV was read or written in V1-2.1. The current integrity of the live DEV object is therefore **not demonstrated**; it is checked when the owner applies M4 to M6 on DEV (contract §16.11). DEV path: `up upgrade 20261028100000` from a copy whose migrations stop at M3, the fixture placed as the operator tool placed the drawing on DEV (service-key upload, pre-M4 confirmation RPC), then M4+M5 by `supabase migration up`. - `storage.objects` row (id, owner, metadata, version, timestamps), the asset row, the historical sheet row and its version (content md5, updated_at): identical before and after. - Changed, as M4/M5 intend: bucket limit 20 971 520 → 2 097 152; service_role privileges ALL → none. - Bytes through the Storage API before and after: identical to the Git file. - After M4+M5, through the Storage API with real tokens: an editor at aal2 reads the 132 997 bytes (sha256 equal to the key); the same editor at aal1, a catalogue reader and an anonymous caller are refused (the historical sheet has no published version). The rehearsal is reproducible from the repository, re-run on `7eca441`, `5620399` and `2379d28` (final decisions pass) with the same results, the only diff lines being the bucket limit and the service_role privileges: `bash .claude/audits/2026-10-28-v1-2-1-m7/m7-rehearsal.sh [work dir]` (sandbox only; it applies M4+M5 with `supabase migration up --db-url` on the sandbox port, a step `scripts/db-sandbox.sh` has no subcommand for) and `m7-read.mjs` (the reads with real tokens).