pi-aia-asf 0.13.2 → 0.13.4

package/CHANGELOG.md

@@ -7,6 +7,95 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.13.4] - 2026-10-16
+
+### Fixed
+
+- **The model-blind-spot research (D1/D2) could not be verified against its own
+  sources, and now it can.** Both deliverables existed; neither had been audited,
+  and the audit found three real problems.
+  - **Seven quotations attributed to unnamed sources were removed.** The document
+    leaned on block quotes from "a soft-delete design guide", "the tenancy
+    source", "a team that learned it the expensive way" and the BEA book — none
+    carrying a URL. Checked by exact-phrase search and by fetching the cited
+    URLs: `orphans-by-policy`, `humans forget filters … RLS policies forget
+    nothing` and `we paid for the trash folder twice` return **zero results**,
+    and the InfoQ and O'Reilly fitness-function URLs return **404**. An
+    unfindable quotation is worse than no citation — it reads as evidence and
+    invites downstream reuse — so each was deleted, and the claims that still
+    matter were restated as explicitly-labelled **T3** reasoning. Where a primary
+    source says it better, it replaced the paraphrase: the PostgreSQL manual's own
+    default-deny text is stronger and checkable than the quote it replaced.
+    The deletions are listed in a withdrawn table rather than dropped silently.
+  - **What remains is now genuinely T1** — fetched and read this pass, with URLs
+    and page numbers: Vernon's four aggregate rules (ArchiLab, ch. 10 pp.
+    347–388), the OWASP IDOR methodology, PostgreSQL §5.9, the Thoughtworks
+    fitness-function article. `SOURCES.md` gained the Model section with T1/T2/T3
+    classes per 02-research.md, including a row that admits the Rails `dependent:`
+    option list was **not** read and must be checked before any option name
+    reaches normative text.
+  - **The inventory had gone stale.** It was measured against `06a–06i` and never
+    re-derived after `06j`, `06k`, the gate layer and the model fitness level
+    shipped — so it still claimed all 14 classes were invisible. Re-scored against
+    the current gate set with the H/M/S enforcement classes: six classes (3
+    concurrency, 5 time, 6 numeric/unit, 7 locale, 9 unit-of-work, 12 environment)
+    remain **completely uninspected**, and that is now stated as a count instead of
+    a mood.
+  - **The audit covered only half the mandate.** The spec says "ASF **and
+    vigilant**"; vigilant had never been audited. It now has its own table — Loop
+    Guardian, spec verification, completeness guard, obligations container,
+    tool-timeout abort — and the honest conclusion: vigilant inspects the agent and
+    the claim, not the deliverable's design, and by the package boundary it must.
+    So for those six classes, **if ASF does not inspect them, nobody does.** There
+    is no second line of defence in the other package.
+  - **The meta-finding got a status.** "Not one gate asserts the non-effect" was
+    written before 06g §3 added non-effect assertions; it now reports itself as
+    half-closed — the language exists, the compulsion exists only where a blocker
+    was built.
+- **`test/test-model-research.mjs` (42 checks) guards all of the above**, because
+  the failure mode is recurrence rather than novelty: it asserts the inventory is
+  contiguous and fully re-scored, that vigilant is audited, that every sourced
+  subsection resolves to a URL or is labelled T3, and that each withdrawn phrase
+  survives only as a record and never as a quotation. **Negative control verified**
+  — injecting an unattributed quote back into §2.6 makes it fail, and removing it
+  passes.
+
+### Notes
+
+- **`npm audit` triage for this line: accepted, not fixed — and the reason is
+  recorded rather than skipped.** Three advisories report (1 moderate,
+  2 high), all tracing to `undici` and `brace-expansion`. None of them ships
+  with ASF: `dependencies` is `{}` and the packed artifact (50 files) contains
+  no `node_modules` and neither package. They are reachable only through the
+  **peer** host package `@earendil-works/pi-coding-agent`, which is the user's
+  own pi installation and is pinned in `devDependencies` solely so typecheck can
+  run against a known host version (`host types 1.2.0`).
+  - `npm audit fix` would bump that pin. It was deliberately **not** run: it
+    would move the host-types version the entire typecheck and suite are
+    calibrated against, for no change whatsoever in what an installed ASF
+    contains. The fix belongs to the pi release, not to this manifest.
+  - Consequence for anyone reading this: an `npm audit` hit on ASF is a signal
+    about your **pi install**, and should be resolved by updating pi. Checking
+    the shipped surface is the `dependencies: {}` line plus the pack list above,
+    not the audit count.
+
+## [0.13.3] - 2026-10-16
+
+### Fixed
+
+- **06j §5 had drifted back into enumerating products.** The section now carries
+  the store-agnostic *rule* — the defect is not SQL-shaped, and the distinction
+  that makes a hand-written cascade decidable is the **reference** (a row removing
+  itself vs one entity deciding the fate of another) — instead of a table of
+  store families and a list of ORM dialects. Which dialects the scanner
+  recognises is an implementation detail in `lib/code-health/model.mjs`, reported
+  per finding so you can see what it looked at; the standard states categories and
+  leaves the rest to `model.extraCascadePatterns` /
+  `model.extraHandWrittenPatterns`. A recognised dialect is not a coverage
+  promise: an unknown persistence shape is a `scanned: 0` you say out loud, not a
+  clean bill of health. The fixtures keep naming the concrete dialects — that is
+  where a pattern set belongs.
+
 ## [0.13.2] - 2026-10-16
 
 ### Added

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-aia-asf",
-  "version": "0.13.2",
+  "version": "0.13.4",
   "description": "Ai Applied Agentic Software Factory — codifies the full software development flow: intake, research, spec capture, adversarial analysis, planning with approval gates, test-first implementation, and release. Requires pi-vigilant, pi-smart-web-search, pi-smart-fetch, and pi-aia-browser.",
   "keywords": [
     "pi-package",

package/skills/aia-asf/references/06j-data-model.md

@@ -179,36 +179,33 @@ requirement: "is this ownership correct?" is judgment and stays in review, while
 "does this file declare a cascade with no `asf-model:` tag?" is measurable.
 
 **It is a tripwire, not an authority — and it is store-agnostic by design.** The
-defect is not SQL-shaped: an instance inheriting its parent's lifecycle happens
+defect is not SQL-shaped. An instance inheriting its parent's lifecycle happens
 in a relational FK action, in an ORM mapping attribute, in a document store's
 delete dependency, in a key-value prefix wipe, or in a loop someone wrote to
 delete children by hand. Three rules run:
 `undeclared-cascade` (a cascade declaration with no ownership declaration),
 `delete-then-reinsert` (rebuild-by-delete over one table or collection), and
-`handwritten-cascade` — the store-agnostic half, for stores that have **no
-constraint to declare the cascade in the first place**:
-
-| Store family | What the cascade looks like when nobody wrote it down |
-|---|---|
-| **MongoDB / Mongoose** | `deleteMany({ userId })` over a child collection — no FK action exists, so this code *is* the cascade |
-| **Firestore** | a recursive `collectionGroup` / subcollection delete; missing a nested subcollection is the standard failure |
-| **DynamoDB** | `DeleteItem` / `BatchWriteItem` driven by a parent key across tables; GSI/LSI are wiped with the table, items are not |
-| **CouchDB** | documents selected and marked `_deleted: true` in one `bulkDocs` |
-| **Redis** | a key-pattern wipe — `keys('session:<id>:*')` then `del`/`unlink`; the prefix *is* the parent |
-| **Elasticsearch** | `deleteByQuery` on a parent field; no referential integrity exists to fall back on |
-
-ORM-level declarations are matched too — `CascadeType.REMOVE`/`ALL`,
-`orphanRemoval = true`, `DeleteBehavior.Cascade`, `onDelete: Cascade`
-(Prisma/Sequelize), `on_delete=models.CASCADE` (Django),
-`dependent: :destroy` / `:delete_all` (Rails), `cascade = "…delete…"`
-(SQLAlchemy) — so the gate is not SQL DDL only.
-
-A fire here asks for a **declaration**; it does not claim the ownership is
-wrong, which stays a judgement in review. A delete of a row by its own identity
-(`deleteOne({ _id: id })`) is not a cascade and does not fire — a rule that
-fires on ordinary deletes gets switched off, and the coverage goes with it.
-Projects add dialects with `model.extraCascadePatterns` and
-`model.extraHandWrittenPatterns` rather than waiting for a release.
+`handwritten-cascade` — the shape a cascade takes in a store that has **no
+constraint to declare it in**, where the decision exists only as code someone
+wrote: a multi-delete driven by a reference to *another* entity.
+
+The distinction that makes `handwritten-cascade` decidable rather than
+judgemental is the **reference**. `deleteOne({ _id: id })` is a row removing
+itself; `deleteMany({ userId })` is one entity deciding the fate of another. The
+first does not fire — a rule that fires on ordinary deletes gets switched off and
+the coverage goes with it, which is why the near-miss fixtures matter more than
+the fire fixtures. A fire asks for a **declaration**; it does not claim the
+ownership is wrong, and that judgement stays in review.
+
+**The pattern set is illustrative and the standard does not enumerate products.**
+The categories above are the standard; the dialects the scanner happens to
+recognise are an implementation detail in `lib/code-health/model.mjs`, reported
+per finding so you can see what it looked at. A project whose store is not in
+that set adds its own spelling with `model.extraCascadePatterns` /
+`model.extraHandWrittenPatterns` rather than waiting for a release — and a
+recognised dialect is not a coverage promise. If your persistence layer speaks a
+shape the scanner does not know, that is a `scanned: 0` you must say out loud,
+not a clean bill of health.
 
 **Where the store has no constraint to inspect, the scan has nothing to attach
 to and the invariant tests below are the only hard gate** — say which of the two