fix(scripts): make the tenant-audit census say when it could not place a receiver's type

[os-elon-musk]

Fixes #18548
Clause-②: no

The defect, in one sentence

scripts/tenant-audit-census.mjs builds its engine type index from git ls-files
and nothing else, so a receiver typed with a declaration that enumeration cannot
see — untracked, generated, or a dependency's — has no type the index can find.
That site resolved to kind: 'other', was subtracted from the certified
population, and printed nothing. It was spelled exactly like "I read the
declared type and it is a Set": same row, same exit 0, no diagnostic.

⛔ The enumeration is unchanged and stays tracked-only. Acceptance item 3 is
explicit that reading untracked files would make the repository's verdict a
function of somebody's working tree, and this PR does not do it. Acceptance
item 4 is what landed instead: the criterion is now the script's own, written
down at CORPUS_TYPE_DECL, and pinned in both directions.

Acceptance item 1 — the reading, re-measured on this tree

origin/main = 02bdeaaf2. 129 write calls reach kind: 'other' through a
declared type or a constructed value; 17 more arrive through an
UNTYPED_RECEIVERS row. 116 of the 146 subtractions are made because the
receiver's declared type is not in the engine type index — the path the card
asks about. Broken out by the fact each one now rests on:

why the subtraction stands	sites
constructed-locally — the census watched new X being constructed	92
ledger-row — an UNTYPED_RECEIVERS row says what the receiver is	17
corpus-type — a type THIS corpus declares; the door rule read it and said no	15
builtin-import — the receiver is imported from a node: builtin	13
platform-type — a language global this corpus cannot declare	6
⚠️ anonymous-type — an inline type literal: no name for the index to be keyed on	2
⚠️ type-not-in-corpus — no declaration of that name exists in the tracked corpus	1

⇒ 3 undefended subtractions on a clean tree, and 2 of them are provably
wrong: their own inline type text states an ObjectQL write door
(insert / update with a first parameter named object or name and typed
string), which is the census's own definition of an engine, applied to named
declarations only:

• packages/plugins/plugin-auth/src/audience-gate-test-support.ts:72 —
  engine.insert('sys_invitation', row, { context: { isSystem: true } })
• packages/plugins/plugin-auth/src/sso-client-secret.ts:303 —
  e.update(SSO_PROVIDER_OBJECT, patch, { context: { isSystem: true } })

Both are real engine writes, and both are subtracted from the population today.
⛔ This PR does not re-place them — that moves a security-relevant population
other cards cite, and acceptance item 2 says diagnostics before behaviour. It
makes them loud and names them for a follow-up card.

The firing control (the card's git add experiment, reproduced)

packages/services/service-settings/src/settings-service.types.ts declares
SettingsEngine. Untracking it with git rm --cached leaves the bytes on disk
untouched (blob 90c0966 before and after) and removes only the declaration the
index can see. Measured four times on this tree:

census	tracked	declaration untracked
pre-change (02bdeaaf2)	225 sites / 146 subtracted, exit 0, stderr 0 bytes	223 sites / 148 subtracted, exit 0, stderr 0 bytes
this PR	225 / 146, exit 0, 3 named warnings	223 / 148, exit 0, 6 named warnings, type-not-in-corpus 1 to 4

The four names the diagnostic prints under the control are SettingsEngine (2
sites), SettingsSecretStore (1) and Hono (1). Pre-change, the same run wrote
zero bytes to stderr in both columns — which is the defect the card describes,
measured rather than relayed.

What changed

• corpusTypeNames() — every type NAME the tracked corpus declares, collected
  from the same git ls-files enumeration as the index, so the diagnostic can
  never see more of the tree than the index it reports on. 6520 files, 103 MB,
  8154 names, 0.6 s. Zero names refuses.
• NON_ENGINE_REASONS — a closed set of seven. Five name a fact that defends
  the subtraction; two admit the census could not place the receiver's type.
• typeTextDeclaresEngineDoor() reuses memberIsEngineDoor, so "an engine door
  in its own text" is one rule with one implementation, not two spellings.
• The undefended ones are printed per site on every run
  (::warning::[receiver-type-not-placed]), carried per site in --json, and
  counted under enforcement in both artefacts with the rows in the audit
  ledger. A type that leaves the index now lands in the diff by name.
• The reason breakdown deliberately stays OUT of the enforced artefacts: every
  commit that adds a Map moves it, which is the ambient churn with no safety
  content the maintainer's 2026-08-31 enforced/unenforced ruling exists to keep
  out. The undefended rows move only when the population's own boundary moves, so
  they belong on the enforced side by that ruling's own criterion.

⚠️ Nothing got quieter and nothing moved. 225 sites, 146 subtracted, 0
unresolved receivers, generator exit 0 and gate exit 0 — before and after, on a
clean tree. No subtraction is withdrawn, no verdict is recomputed, no exit code
changed. The only committed-number movement is in the two unenforced rows a
regeneration always touches: tracked non-test sources scanned 570 to 573 and
the dated marker 2026-09-16 at 11daf7f69 to 2026-09-18 at 02bdeaaf2.

⚠️ The generator warns rather than refuses, deliberately: this class is not empty
on a clean tree, so refusing would red main for findings nobody has ruled on,
and a gate that reds on arrival is a gate that gets weakened.

Acceptance item 4 — the criterion, pinned in both directions

The census self-test goes 30 to 44 cases. The load-bearing pair is one receiver,
one source text, and index membership as the only variable:

• type IS in the index ⇒ engine/IProbeEngine
• type is NOT in the index ⇒ other/type-not-in-corpus, naming IProbeEngine

Both arms are also pinned at the classifier: a name the corpus declares reads
corpus-type; the same name absent from the corpus reads type-not-in-corpus.
The five defensible arms are pinned too, so the diagnostic cannot decay into
"everything is undefended" — a pile that flags all 146 hides the 3 that matter
exactly as effectively as flagging none.

Both ablations fired (scripts/ablation-replace.mjs, on-disk blob evidence,
restore verified against HEAD):

mutation	result
const unseen = ids.filter(...) to const unseen = [] (collapse the tracked-only distinction)	✗ 2 of 44 — exactly the two criterion pins
door = true to door = false in the door reader	✗ 1 of 44 — the door pin

Restored: blob back to b87aade77de1, git diff HEAD empty, 44 of 44 green.

Premises tested

1. The relayed headline reproduces exactly. The card relays 223 / 146 to 148 versus 225 / 146 from card [finding] tenant-audit 普查器的 ERROR 方向送不到 CI —— 门禁一次都没读 unledgered / staleLedgerRows #18211's dev and states it did not reproduce
   them. They reproduce on this tree, to the digit, with the file named above.
2. The enumeration claim holds, with one correction. On 02bdeaaf2:
   buildEngineTypeIndex is at :215, trackedTs at :163, the ls-files
   call at :164, and grep -c "ls-files" is 1 — one mouth, no second path.
   ⚠️ But trackedTs( appears 5 times: one definition and four
   consumers (collectSources:171, buildEngineTypeIndex:217,
   widenIndexThroughAliases:313, declaredObjects:638). The card's control
   proves there is no second way in; it does not show that four separate answers
   are derived from that one way. This PR adds a fifth consumer on purpose, for
   exactly that reason.
3. It is subtraction, not an under-reported row. Under the control the
   population itself moves 225 to 223 and object name chosen at run time moves
   76 to 74: the two sites leave census.totals entirely rather than being
   counted and mis-verdicted. Confirmed structurally too — kind === 'other'
   increments nonEngineCalls and never pushes into sites.
4. origin/main merged before opening this PR: already up to date at
   02bdeaaf2, and every reading above is from this tree.

Acceptance notes

• Noted, not filed — the two provably wrong subtractions. The two
  anonymous-type sites above satisfy the census's own door rule and are
  subtracted anyway, because the rule is keyed on named declarations while these
  types are spelled inline. Re-placing them moves the certified population 225 to
  227 and rewrites both artefacts plus the page's hand-written deviation prose —
  a numbers change other cards cite, which wants its own card and its own review,
  not a rider on a diagnostics card. They are now named on every run and counted
  under enforcement, so they are no longer invisible. Carrier: the seat that
  files it.
• Same-mechanism cards stay out. [finding] check-bash32-floor's census shrinks silently on a partial checkout — listPopulation lists from the git INDEX and reads from DISK, skipping unreadable paths at exit 0 with no diagnostic #18465 (check-bash32-floor.mjs's
  listPopulation shrinking on a partial checkout) is the nearest relative and is
  not touched here. If it helps whoever takes it: the shared shape is
  "an index-derived population shrinks silently", and the repair shape this PR
  uses is make the subtraction state the fact it rests on, and enforce the
  residue that states none — the population stays where it is and the silence is
  what goes away. [regression] check:cross-package-test-inputs 的裁决取决于被 gitignore 的空目录 packages/spec/dist 存不存在 —— 构建过的工作树一律红,CI 绿只因那一步跑在构建之前(#18340 引入) #18348 is likewise untouched.
• The docs/audits/ ledger and the page were regenerated with
  node scripts/tenant-audit-census.mjs --write, the one repair path.

维护者速读(草稿)

改了什么 —— tenant-audit 普查器在把一个写调用点判成「非引擎调用」并从被认证总体里减掉时,
现在必须说出它凭的是哪条事实。七个理由构成一个封闭集:五个是站得住的(node: 内建、就地
new 构造、语言全局、本仓声明且门规则读过后否掉的类型、台账行),两个是坦白「我没能把这个接
收者的类型放到位」。⛔ 枚举口没动,仍然只从 git ls-files 建 —— 这是设计,不是缺陷。

为什么改 —— 原来这两类印出来完全同形:总体静默变小、exit 0、零诊断。本席实测:把一个声明
引擎接口的文件 git rm --cached(磁盘字节不变),改前的普查器给出 223 个站点(原本 225)、
减法 148(原本 146),而 stderr 一个字节都没有。⇒ 这个仪器的裁决确实取决于你有没有先跑
git add,而它对此一言不发。

风险与代价(含回滚) —— 本次没有任何判定、计数或退出码移动(225/146/0,改前改后一致),
没有撤回任何一次减法。代价是两份产物各多一节「无法辩护的减法」并进入强制比对:今后一个类型离开
索引会让门禁红着报出该类型的名字,而不是安静少算一个站点。回滚 = 单 commit revert 再跑一次
node scripts/tenant-audit-census.mjs --write,无数据迁移、无发布面。

席位意见 ——

你要做的 —— 无需动作;这是仓内普查脚本,不发布任何东西(声明行在正文首部)。⚠️ 若关心那两个
「自己的类型文本就写着引擎门、却仍被减掉」的站点(audience-gate-test-support.ts:72、
sso-client-secret.ts:303),它们会独立立卡:把它们放回总体会把 225 推到 227 并改写页面里手写
的偏差表,该由一次专门的评审决定。

🤖 Generated with Claude Code

https://claude.ai/code/session_019srGWGCBBCBHqcDoRZpQRh

---

Generated by Claude Code