fix(spec): 不再承诺根本不会跑的校验，并让严格性账本闸门看见嵌套目录（#4001） (#4412)

* fix(spec): stop promising validation that never runs, and let the strictness ledger see nested dirs (#4001)

Two corrections to the #4001 unknown-key campaign. Both are the same defect the
campaign exists to remove — a success signal covering an omission — found in the
campaign's own output.

1. The `config` prescription was false.

`DatasourceSchema` went strict in #4207 with `config` left open, and the
rejection for a misplaced connection key prescribed: "Move it to
`config: { host: … }`; the driver's own configSchema validates it there."

Nothing validates it there. `DriverDefinitionSchema.configSchema` is a
`z.record`; both bundled driver specs set it to `{}` (mongo's with a comment
promising runtime population that no code performs); and no consumer in this
repo reads it. So the message took an author who had made a recoverable mistake
at a place that now catches it and directed them, with the platform's authority,
at a slot where the same mistake is silent again — `config: { hostname: … }` is
stripped and the datasource connects on localhost, which is #4001's original bug
one level down.

The guidance now names the per-driver shape to write against
(`PostgresConfigSchema` / `MongoConfigSchema` / `MemoryConfigSchema`) and claims
nothing about enforcement. Same false claim removed from `driver/mongo.zod.ts`,
whose header advertised that the Platform validates `datasource.config` with it.
Enforcement — enforce-or-remove on an inert `configSchema` — is filed as #4410.

Error-message and documentation text only; no authorable key changed.

2. The ledger gate's coverage walk was one level deep.

`check:strictness-ledger` promises every sited `*.zod.ts` in a triaged directory
carries a classification row. It listed directories non-recursively, so
`data/driver/` — three per-driver config files, nine authorable sites — was
invisible to the check whose entire promise is "no undeclared surface", which
printed "no undeclared schema files" and was believed.

A gate that under-reports is worse than no gate: it converts "I should classify
this" into "it is already classified". The walk is now recursive, the three
files are classified, and the ledger records both findings.

`scripts/strictness-ledger.test.ts` pins the recursion, and is not redundant with
the gate: measured, with the rows now present, reverting to a non-recursive walk
leaves the gate GREEN — it cannot catch its own regression. The test can, and
was verified to go red before it went green.

Verified: spec 283 files / 7205 tests, `tsc --noEmit` clean, all 8 generated
artifacts current, and all 10 `check:*` gates green.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WnqGjQFQMqd5k81LYV8SCY

* fix(docs): v17 release notes repeated the false driver-configSchema claim (#4001)

Fourth and last copy, and the only user-facing one. The v17 notes told readers
that `config` and `readReplicas` stay open records because "the driver's own
`configSchema` validates them" — the same sentence corrected in
`datasource.zod.ts` (module comment + the `belongsInConfig` rejection) and
`driver/mongo.zod.ts`.

Found by this PR's own docs-drift advisory, which is worth recording: the bot
listed 107 docs on a package-level heuristic and the correction was real in
exactly one of them. Cheap to check, and it was the copy readers actually see.

Now states that nothing validates inside `config`, that the per-driver schemas
exist but parse nothing, and that a misspelling one level down is still silent —
with the pointer to #4410.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WnqGjQFQMqd5k81LYV8SCY

---------

Co-authored-by: Claude <noreply@anthropic.com>

Author: os-zhuang

.changeset/strictness-ledger-recursive-coverage.md

@@ -0,0 +1,13 @@
+---
+'@objectstack/spec': patch
+---
+
+Datasource unknown-key guidance no longer promises validation that does not happen, and the #4001 strictness ledger sees nested schema directories.
+
+Two related corrections to the #4001 unknown-key campaign, both about a check reporting more coverage than it had.
+
+**The `config` prescription was false.** When `DatasourceSchema` rejects a connection detail written at the top level, it prescribed: "Move it to `config: { host: … }`; the driver's own configSchema validates it there." Nothing validates it there — `DriverDefinitionSchema.configSchema` is a `z.record` that both bundled driver specs set to `{}`, and no consumer reads it. The message therefore took an author who had made a recoverable mistake at a place that now catches it and pointed them at a slot where the same mistake is silent again: `config: { hostname: … }` is dropped and the datasource connects on localhost. The guidance now names the per-driver shape to write against (`PostgresConfigSchema` / `MongoConfigSchema` / `MemoryConfigSchema`) instead of promising a gate. Enforcement is tracked in #4410. The same false claim is removed from `data/driver/mongo.zod.ts`, whose header advertised that the Platform validates `datasource.config` with it.
+
+No authorable key changed — this is error-message and documentation text only.
+
+**The ledger gate's coverage walk was one level deep.** `check:strictness-ledger` promises that every `*.zod.ts` with sites in a triaged directory carries a classification row. It listed each directory non-recursively, so `data/driver/` — three per-driver config files, nine authorable sites — was invisible to it while it printed "no undeclared schema files". The walk is now recursive (nested files declare as `driver/postgres.zod.ts`), those three files are classified in the ledger, and `scripts/strictness-ledger.test.ts` pins the recursion — necessary because with the rows in place the gate itself passes either way and cannot catch its own regression.

content/docs/releases/v17.mdx

@@ -412,8 +412,12 @@ schema to the two highest-risk authorable surfaces, per the triage in
 - **Datasources** — `DatasourceSchema` with its `pool` / `healthCheck` / `ssl` /
   `retryPolicy` blocks, the ADR-0015 `external` federation settings and their
   `validation` policy, `DatasourceCapabilities`, and `DriverDefinitionSchema`.
-  `config` and `readReplicas` stay **open** records: their shape is per-driver
-  and the driver's own `configSchema` validates them. That openness is why the
+  `config` and `readReplicas` stay **open** records: their shape is per-driver.
+  Nothing validates *inside* them — an earlier version of this note said the
+  driver's own `configSchema` did, which was wrong; the per-driver schemas exist
+  (`PostgresConfigSchema` and siblings) but nothing parses `config` against
+  them, tracked as #4410. So a misspelling one level *down* is still silent
+  today. That openness is why the
   top level had to close — a connection key written one level too high (`host`
   next to `driver` instead of inside `config`) was stripped, and the datasource
   then connected on driver defaults rather than failing. Those keys now

docs/audits/2026-07-unknown-key-strictness-ledger.md

@@ -101,6 +101,18 @@ dropped at parse, and nothing failed.
    now aliases to `expanded`. Note where this one was found: not in a tenant
    project, but in first-party platform metadata that had been shipping for
    releases.
+7. **This campaign's own fix signposted the way into the failure mode it
+   exists to kill.** The strict rejection on a misplaced `host` prescribed:
+   "Move it to `config: { host: … }`; the driver's own configSchema validates
+   it there." False twice over — `DriverDefinitionSchema.configSchema` is a
+   `z.record` that both bundled driver specs set to `{}`, and nothing in the
+   repo reads it. So an author who made a *recoverable* mistake at a place that
+   now catches it was directed, with the platform's authority, at a slot where
+   the same mistake is silent again: `config: { hostname: … }` is stripped and
+   the datasource connects on localhost — #4001's original bug verbatim, one
+   level down. Corrected to name the per-driver shape instead of promising a
+   gate; enforcement is #4410. **A wrong instruction is worse than none**, and
+   worst for an AI author, whose only signal is whether the parse complained.
 
 This is the empirical argument for the ratchet: the inference "no metadata in
 the repo carries unknown keys" was **false three times over**, and only the
@@ -146,7 +158,7 @@ tightening (the #4001 "sharing-rule lesson": candidates, not verdicts).
 | `notification.zod.ts` / `offline.zod.ts` / `report.zod.ts` | 3 ea | authorable (p) | |
 | `sharing.zod.ts` | 2 | authorable (p) | public-sharing config |
 
-### `data/` — 149 sites
+### `data/` — 158 sites
 
 | File | Sites | Class | Note |
 |---|---|---|---|
@@ -157,7 +169,8 @@ tightening (the #4001 "sharing-rule lesson": candidates, not verdicts).
 | `field.zod.ts` | 11 | authorable | partially strict |
 | `filter.zod.ts` / `query.zod.ts` | 11+5 | open | query dialect — user data flows through; validated semantically elsewhere. `query.zod.ts` dropped one site in #4196: `FieldNodeSchema`'s nested-select object form was declared-but-inert and narrowed to `z.string()`, so the union's second member is gone. Four more left in #4286 with the `joins`/`windowFunctions` removals: `JoinNodeBaseSchema`, `WindowFunctionNodeSchema`, and `WindowSpecSchema`'s two blocks (outer + `frame`) were deleted with their clusters. Class unchanged |
 | `driver-nosql.zod.ts` / `driver.zod.ts` / `driver-sql.zod.ts` | 10+9+2 | wire | driver capability contracts |
-| `datasource.zod.ts` | 9 | authorable | **strict as of #4001 data step** — all 9: `DatasourceSchema` (+ `pool` / `healthCheck` / `ssl` / `retryPolicy`), `ExternalDatasourceSettingsSchema` (+ `validation`), `DatasourceCapabilities`, `DriverDefinitionSchema`. `config` + `readReplicas` stay `z.record` by construction (per-driver shapes; the driver's own `configSchema` validates them) — which is precisely why the top level had to close: a connection key written one level too high was stripped, and the datasource then connected on driver defaults instead of failing |
+| `datasource.zod.ts` | 9 | authorable | **strict as of #4001 data step** — all 9: `DatasourceSchema` (+ `pool` / `healthCheck` / `ssl` / `retryPolicy`), `ExternalDatasourceSettingsSchema` (+ `validation`), `DatasourceCapabilities`, `DriverDefinitionSchema`. `config` + `readReplicas` stay `z.record` by construction (per-driver shapes — see the `driver/` row below). This row used to add "the driver's own `configSchema` validates them"; **it does not, and never did** — corrected, and the gap is #4410. Which is precisely why the top level had to close: a connection key written one level too high was stripped, and the datasource then connected on driver defaults instead of failing |
+| `driver/memory.zod.ts` / `driver/mongo.zod.ts` / `driver/postgres.zod.ts` | 6+1+2 | authorable | The per-driver shapes for the `config` slot — what an author actually writes under `datasource.config` (`host`, `port`, `filename`, pool sizes). **Undeclared here until the coverage walk went recursive** (see below): a subdirectory was invisible to the gate, so these nine sites sat outside the map while the map reported full coverage. Authorable by the rule, but they are **contract-only exports today** — nothing parses `datasource.config` against them and both `*DriverSpec.configSchema` literals are `{}` (#4410). Strictness here would therefore enforce nothing; this row is blocked on #4410 giving it a parse site, not on a verification pass |
 | `analytics.zod.ts` | 8 | mixed (p) | |
 | `document.zod.ts` | 8 | wire (p) | |
 | `hook.zod.ts` / `hook-body.zod.ts` | 6+2 | mixed | **strict as of #4001 data step** for the AUTHORING shapes: `HookSchema` (+ `retryPolicy`) and both body branches (`ExpressionBodySchema` / `ScriptBodySchema`). `HookContextSchema` and its `session` / `provenance` / `user` blocks are the RUNTIME shape the engine hands a handler — they stay tolerant, and must: strictness there would make an engine-internal enrichment (as `provenance` was in #3712) a breaking change for anyone parsing a context they were given. The file's old blanket `authorable (p)` was too wide — verification split it |
@@ -220,6 +233,32 @@ tightening (the #4001 "sharing-rule lesson": candidates, not verdicts).
    strict close-out on what it actually reports — which is the whole point of
    having built it. Nothing more to do here until there is field data.
 
+   **This wait has a decision point, deliberately.** "Wait for field data" with
+   no way to tell when it has arrived is how a ratchet stops without anyone
+   choosing to stop it — and this file would go on describing an in-flight
+   campaign either way. So the wait is discharged by an answerable question, not
+   by a date: *has `lintUnknownAuthoringKeys` reported an unknown key on any
+   surface outside this repo yet?* Three outcomes, each with a next action:
+   - **Findings exist** → they are the close-out worklist. Tighten the shapes
+     they name first; that is the evidence the whole layer was built to produce.
+   - **Zero findings, and the layer is reaching real authors** → the tail is
+     cheaper than feared and the remaining directories can be batched by class
+     rather than one shape at a time.
+   - **Zero findings because nothing is reporting back** → then the layer is not
+     instrumented, and *that* is the next task, not more strictness. This is the
+     outcome to actually check for: it is indistinguishable from success at a
+     glance, which is this campaign's own subject matter.
+
+   Whoever reads this next: answer the question and record the answer here, even
+   if the answer is "still nothing". A wait that is never re-examined is
+   indistinguishable from an abandoned one.
+
+2. `studio/` is the largest untouched authorable block — 27 sites, **0 strict**,
+   and all three files still carry a provisional `(p)` from the original triage.
+   Not blocked on field data (Studio-written JSON is our own producer, so the
+   downstream risk is the lowest on the board); it is simply unstarted. If the
+   step-1 question comes back "nothing is reporting", start here instead.
+
 Done in step 2: `security/rls.zod.ts` + `security/sharing.zod.ts` strict;
 `PositionSchema` strict with the protection envelope declared (closing the
 known sibling gap below).
@@ -303,9 +342,11 @@ checkable, so this map cannot go stale in silence again:
   matches means schemas were added or removed under a `Class` verdict nobody
   re-examined. Touching a file forces you back through this ledger.
 - **Coverage.** Every `*.zod.ts` in a triaged directory that HAS sites must have
-  a row. A new one is undeclared surface. Zero-site files (pure enum/token
-  modules like `data/date-macros.zod.ts`) are skipped — there is nothing to
-  classify — and become reportable the day they grow their first `z.object(`.
+  a row. A new one is undeclared surface. The walk is **recursive**; nested files
+  are declared by their path relative to the section directory
+  (`driver/postgres.zod.ts`). Zero-site files (pure enum/token modules like
+  `data/date-macros.zod.ts`) are skipped — there is nothing to classify — and
+  become reportable the day they grow their first `z.object(`.
 - **Section totals**, and that any row claiming "strict as of" names a file that
   really contains `.strict()`.
 
@@ -336,5 +377,23 @@ precisely the intended behaviour: the file arrived, so someone had to classify
 it. It is now a row. Every existing count survived that merge unchanged, so the
 failure was exactly as narrow as it should have been.
 
+**And then the gate turned out to have the ledger's own disease.** Its coverage
+walk listed each triaged directory exactly one level deep, so `data/driver/` —
+three per-driver connection-config files, nine sites — was invisible to the check
+whose entire promise is "no undeclared surface". The gate printed *"no undeclared
+schema files"* while nine authorable sites sat outside the map. Fixed by making
+the walk recursive; the three files are now a row.
+
+Read that next to this file's own opening argument — *a map that drifts is worse
+than no map, because it is followed*. The same asymmetry applies one level up,
+and harder: **a gate that under-reports is worse than no gate, because it
+converts "I should classify this" into "it is already classified."** No gate
+leaves a reader suspicious; a green gate retires their suspicion. That is the
+identical shape to the silent strip this whole campaign is about — a success
+signal covering an omission — reproduced in the instrument built to detect it.
+So when a check claims coverage, prove it sees something it is supposed to see
+before trusting the green: this one was verified by watching it go red on
+`data/driver/` and green again only once the rows existed.
+
 Long tail stays gated on a verification pass per shape — never a one-shot
-"make all ~453 sites strict" (ADR-0054 ratchet; #4001's own recommendation).
+"make all ~500 sites strict" (ADR-0054 ratchet; #4001's own recommendation).

packages/spec/scripts/check-strictness-ledger.mts

@@ -23,10 +23,14 @@
 //      touching a file forces you back through the ledger.
 //   2. Coverage. Every `*.zod.ts` under a triaged directory that HAS `z.object(` sites
 //      must appear in that directory's table. A new one is undeclared surface —
-//      exactly what the ledger exists to prevent. Files with zero sites (pure enum /
-//      token modules like `data/date-macros.zod.ts`) are skipped: the ledger classifies
-//      sites, and they have none to classify. This is not a hole — the day such a file
-//      grows its first `z.object(` it becomes undeclared and this gate says so.
+//      exactly what the ledger exists to prevent. The walk is RECURSIVE; nested files
+//      are declared by their path relative to the section directory
+//      (`driver/postgres.zod.ts`). It was not recursive at first, and `data/driver/`
+//      sat undeclared behind that — see the ledger's note on it. Files with zero sites
+//      (pure enum / token modules like `data/date-macros.zod.ts`) are skipped: the
+//      ledger classifies sites, and they have none to classify. This is not a hole —
+//      the day such a file grows its first `z.object(` it becomes undeclared and this
+//      gate says so.
 //   3. Section totals. `### \`ui/\` — 192 sites` must equal the sum of its rows.
 //      Cheap, and it catches a row edited without updating the header.
 //   4. Strictness claims. A row whose note says "strict as of" must name a file that
@@ -48,6 +52,8 @@ import fs from 'node:fs';
 import path from 'node:path';
 import url from 'node:url';
 
+import { countSites, listSchemaFiles } from './lib/strictness-ledger';
+
 const HERE = path.dirname(url.fileURLToPath(import.meta.url));
 const SPEC = path.resolve(HERE, '..');
 const REPO = path.resolve(SPEC, '../..');
@@ -79,11 +85,6 @@ function parseCounts(cell: string, fileCount: number): number[] | null {
   return nums.length === fileCount ? nums : null;
 }
 
-/** `z.object(` occurrences — the ledger's own stated counting method. */
-function countSites(file: string): number {
-  return (fs.readFileSync(file, 'utf-8').match(/z\.object\(/g) ?? []).length;
-}
-
 const md = fs.readFileSync(LEDGER, 'utf-8').split('\n');
 const rows: Row[] = [];
 const sectionTotals = new Map<string, { declared: number; line: number }>();
@@ -169,7 +170,10 @@ for (const row of rows) {
 for (const [d, declared] of declaredByDir) {
   const dirPath = path.join(SRC, d);
   if (!fs.existsSync(dirPath)) continue;
-  const onDisk = fs.readdirSync(dirPath).filter((f) => f.endsWith('.zod.ts'));
+  // Recursive — see lib/strictness-ledger.ts for why that is load-bearing.
+  // Nested files are declared by their path relative to the section directory
+  // (`driver/postgres.zod.ts`), which the row parser already accepts.
+  const onDisk = listSchemaFiles(dirPath);
   // Zero-site files carry nothing to classify (see the header note). They become
   // reportable the moment they grow a `z.object(`.
   const missing = onDisk.filter((f) => !declared.has(f) && countSites(path.join(dirPath, f)) > 0);

packages/spec/scripts/lib/strictness-ledger.ts

@@ -0,0 +1,39 @@
+// Copyright (c) 2025 ObjectStack. Licensed under the Apache-2.0 license.
+
+/**
+ * Shared pieces of the #4001 strictness-ledger gate
+ * (`../check-strictness-ledger.mts`), extracted so the gate and its regression
+ * test cannot drift apart on the one property that already broke silently:
+ * whether the coverage walk descends into subdirectories.
+ */
+
+import fs from 'node:fs';
+import path from 'node:path';
+
+/** `z.object(` occurrences — the ledger's own stated counting method. */
+export function countSites(file: string): number {
+  return (fs.readFileSync(file, 'utf-8').match(/z\.object\(/g) ?? []).length;
+}
+
+/**
+ * Every `*.zod.ts` under `dir`, **recursively**, as `/`-separated paths relative
+ * to `dir` (so a nested file reads `driver/postgres.zod.ts` — exactly how the
+ * ledger declares it).
+ *
+ * The recursion is the whole point of this function existing. The gate's first
+ * version listed each triaged directory one level deep, which made
+ * `data/driver/` — three per-driver connection-config files, nine authorable
+ * sites — invisible to the check whose entire promise is "no undeclared
+ * surface". It printed "no undeclared schema files" and was believed.
+ *
+ * A gate that under-reports is worse than no gate: it converts "I should
+ * classify this" into "it is already classified". Keep the walk recursive, and
+ * see `strictness-ledger.test.ts`, which fails if it stops being.
+ */
+export function listSchemaFiles(dir: string): string[] {
+  return fs
+    .readdirSync(dir, { recursive: true })
+    .map((f) => String(f).split(path.sep).join('/'))
+    .filter((f) => f.endsWith('.zod.ts'))
+    .sort();
+}