fix(plugin-security): the boot heal converges on duplicated permission-set names, and a refused existence read never inserts

[objectstack-fleet]

Fixes #22169

Clause-②: no

What was wrong (measured on main at 6ed0c0f3e5)

All three mechanisms the card names were confirmed on current main before any edit, with an in-repo reproduction: a real ObjectQL over a real SqlDriver (better-sqlite3, in memory), a real ObjectStackProtocolImplementation, and the per-organization unique index dropped after schema sync. Dropping it gives the shape of a remote table built before the index retrofit, the only shape that can hold duplicates.

1. The heal did not converge. reconcilePermissionSetProjection handed each drifted, overlay-less record to upsertEnvPermissionSet, which re-read the record by name (first row, lowest id). With 8 declared names × 3 rows, boot 1 wrote 8 rows, logged 24 drift warnings and left 16 rows drifted. Boot 2, boot 3 and every later boot logged 16 warnings and wrote nothing.
2. A layered read was discarded on every row. getMetaItemLayered was awaited for every candidate row before the trust rule decided whether it needed the answer. While a SchemaRegistry is readable, the rule never uses it. One layered read is 4 statements on the real protocol, so the shape above paid 24 × 4 per boot.
3. A refused existence read inserted a duplicate. bootstrapPlatformAdmin reads each default set by name through a tryFind that answers [] for a refused read, and it inserted on both "absent" and "refused". The heal's by-name upsert had the same insert branch. When by-name reads were refused, the unpatched heal inserted 24 rows (24 → 48) and the bootstrap inserted 2 more.

The change (@objectstack/plugin-security only: 2 source files)

• permission-set-projection.ts, reconcilePermissionSetProjection heal branch:
  • The heal writes the row it read, by its id, facets only. It calls tryUpdate with { id: row.id, ...permissionSetRowFields(effective) }, the same columns the upsert's update leg wrote. Identity, active, customized and provenance are not touched. The heal no longer asks an existence question, so it has no insert branch.
  • The drift line is logged after the write. It names the row's id, and a refused write is logged as refused instead of "re-projected". driftHealed counts landed writes, now one per drifted row.
  • The layered read is lazy, once per name. It is asked only when no SchemaRegistry answers, and the answer is cached per name for the pass. The cache entry is dropped after this pass backfills that name, because the backfill changes the answer. A later row of that name then sees the backfilled definition and heals to it, instead of backfilling a second time.
• bootstrap-platform-admin.ts, the default-set seed loop:
  • A refused existence read declines the insert. tryFind's existing onRefusal observer tells "refused" apart from "empty".
  • The pass says so once. It writes one warn line per pass, not one per set, naming the sets. The line goes through reportThroughSink, the derivation the sibling seeders' unreadable lines use. When admin_full_access is one of the declined sets, the promotion half returns admin_permission_set_unreadable instead of admin_permission_set_missing. Nothing in the repository branches on that string; the only reads are doc comments.
• One patch changeset for @objectstack/plugin-security.

Not changed (card scope):

• No existing duplicate is deleted; that is an operator repair.
• upsertEnvPermissionSet's per-item read on the live single-mutation path is unchanged. A refused read there still inserts.
• No change to packages/spec, no new gate.

Measurement: statements per boot, in-repo reproduction

This was measured with the harness above on the fix at 5ccce70966 (source unchanged since; later commits touch tests and the changeset only). Each count is every SQL statement knex emitted during reconcilePermissionSetProjection. The harness was not committed.

shape	boot	main 6ed0c0f3e5	this PR
8 names × 3 rows, all drifted	1	148 statements (140 select / 8 update), 24 layered reads, 24 drift warns, 8 healed, 16 still drifted	76 statements (52 select / 24 update), 0 layered reads, 24 drift warns, 24 healed, 0 drifted
	2, 3	116 statements, 24 layered reads, 16 drift warns, 16 still drifted	4 statements, 0 layered reads, 0 warns
60 names × 5 rows, all drifted	1	1684 (1624 / 60), 300 layered reads, 300 warns, 60 healed, 240 still drifted	904 (604 / 300), 0 layered reads, 300 warns, 300 healed, 0 drifted
	2, 3	1444 statements, 300 layered reads, 240 drift warns	4 statements, 0 layered reads, 0 warns
by-name reads refused, 8 × 3	heal + bootstrap	heal inserted 24, bootstrap inserted 2 (24 → 50 rows)	0 inserts (24 → 24), 1 bootstrap line

On the healing boot, each by-id UPDATE costs the engine one extra SELECT. That is the 52 − 4 = 48 selects for 24 updates.

⛔ The hosted timing is not measured here. The card's cloud reproduction (cloud's ArtifactKernelFactory on the hosted Turso face, ~36 ms per statement, 99.85 s / 170.70 s) cannot be reached from this session, and no cloud time is claimed. The before/after boot times on that reproduction belong to the cloud seat.

Pins (new) and ablations

Pins:

• permission-set-projection-heal-by-id.test.ts (real engine, spied and never replaced; index dropped). Five cases:
  • A duplicated set heals once, by id, and the next boot logs no drift warn.
  • Control: a single drifted row heals as before.
  • No layered read while a registry answers.
  • Without a registry, the layered read happens once per name.
  • A refused by-name read inserts nothing.
• bootstrap-platform-admin-refused-existence-read.test.ts:
  • A refused read inserts nothing, logs one line and reports admin_permission_set_unreadable.
  • Two controls: an empty answer still seeds, and an existing row is left alone.
• permission-set-projection.test.ts gains one case: a just-backfilled name is re-read and the second row heals to the backfilled body, with one saveMetaItem, not two.

Ablations. Each leg ran on e41760d89c through scripts/ablation-replace.mjs in WRAP mode, with the anchor-hit count and blob change verified on disk. Each restore was proven with blob == HEAD and an empty git diff HEAD. The subjects are imported by relative path from src/, so no dist/ leg applies.

leg	mutation	result
A1	heal through the by-name upsert again	4 red: duplicate heal wrote ['ps_a'] instead of ['ps_a','ps_b']; refused read inserted 2; no-registry case ['ps_a']; backfill case driftHealed 0. The control and the no-layered-read case stayed green.
A2	eager per-row layered read restored	3 red: no-layered-read case saw 6 reads instead of []; once-per-name and backfill read counts were also red
A3	per-name cache removed	1 red: ['member_default','member_default'] instead of one read
A4	post-backfill cache drop removed	1 red: 2 saveMetaItem calls instead of 1
A5	bootstrap decline removed	1 red: 3 inserts instead of []; both controls stayed green

A first A2 attempt was a refused no-op: its replacement contained the anchor, so the anchor count stayed x1 → x1, nothing ran, and the restore was proven. It was re-run with a falling anchor.

Verification (HEAD e41760d89c)

• pnpm --filter @objectstack/plugin-security exec vitest run --maxWorkers=2: 176 files, 3695 passed, 45 skipped.
• pnpm --filter @objectstack/plugin-security typecheck exits 0, including check:test-typecheck OK. tsc -p tsconfig.test.json --listFiles lists all 3 touched test files.
• Gates:
  • node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack derived 66 commands, plus pnpm check:durability-log-level; all exit 0.
  • --ran: 66 derived, 66 run, 0 NOT-MEASURED, 0 UNRUN.
  • The first run, at af7b0a339f, had check:engine-double-contract and check:objectql-double-limit red on this PR's new test doubles. Both were fixed in e41760d89c: the heal tests spy on the real engine instead of an object-literal double, and the bootstrap double applies limit by presence after the filter.
• ESLint, narrowed and proven: 5 touched files, 0 errors, 0 warnings, run with --no-inline-config --format json.
  • All 5 files are in the population: --print-config resolves each, and no ignored-file warning was emitted.
  • The repository config never enables type-aware linting (eslint.config.mjs, no parserOptions.project), so this diff cannot move any untouched file's verdict.
• No plugin init()/start() or declaration was edited, so check:init-service-contract and check:startup-registry-verdict were not owed.

Acceptance notes

• Measured, not fixed (scope fence): the overlay pass leaves a duplicated overlay-backed name unconverged, silently.
  • Probe: real engine, 2 drifted rows of member_default plus an active env sys_metadata overlay for it.
  • Boot 1 projected the overlay onto ps_a only, and ps_b kept its drifted facets on boots 1, 2 and 3 with 0 warnings. Step 1 resolves one row per name through buildExistingByName, and step 2 skips every overlay name.
  • It costs no boot time and logs nothing. It is display drift only, because enforcement reads the metadata definition. It is the same by-name family, outside this card's fence.
• Reserved by the card for its own card: upsertEnvPermissionSet's per-item read on the live single-mutation path still inserts on a refused read.
  • retirePermissionSetRecord has the same by-name LIMIT 1 shape: on a metadata delete it removes only the first row of a duplicated name. Read, not exercised.
• feat(core,objectql,plugin-security,plugin-sharing): the catalog is read from the registry; assignment tables reference it by name (ADR-0131 D2/D3/D4) #15196's census marks this machinery for retirement in C3 (refactor(plugin-security,platform-objects,spec): retire the catalog seeders, the per-organization catalog machinery and the four catalog objects; Setup creation is an environment write under single and refused under a wall (ADR-0131 D2/D3/D5/D13) #15204, pm:blocked, not scheduled). C3 would delete these paths. That does not change this card.

---

Generated by Claude Code

[github-actions]

📓 Docs Drift Check

This PR changes 1 package(s): @objectstack/plugin-security, touching 5 documentable anchor(s).

6 hand-written doc(s) NAME something this change touched and may need an implementation-accuracy re-verification:

• content/docs/api/error-catalog.mdx (via sys_permission_set (literal, a string literal in bootstrapPlatformAdmin; a string literal in reconcilePermissionSetProjection))
• content/docs/concepts/metadata-lifecycle.mdx (via sys_permission_set (literal, a string literal in bootstrapPlatformAdmin; a string literal in reconcilePermissionSetProjection))
• content/docs/deployment/environment-variables.mdx (via sys_permission_set (literal, a string literal in bootstrapPlatformAdmin; a string literal in reconcilePermissionSetProjection))
• content/docs/permissions/authorization.mdx (via sys_permission_set (literal, a string literal in bootstrapPlatformAdmin; a string literal in reconcilePermissionSetProjection))
• content/docs/permissions/delegated-administration.mdx (via sys_permission_set (literal, a string literal in bootstrapPlatformAdmin; a string literal in reconcilePermissionSetProjection))
• content/docs/permissions/permission-sets.mdx (via sys_permission_set (literal, a string literal in bootstrapPlatformAdmin; a string literal in reconcilePermissionSetProjection))

⛔ 7 release-owned page(s) also name something this change touched. These are read-only:

• content/docs/releases/v12.mdx (via sys_permission_set (literal, a string literal in bootstrapPlatformAdmin; a string literal in reconcilePermissionSetProjection))
• content/docs/releases/v13.mdx (via sys_permission_set (literal, a string literal in bootstrapPlatformAdmin; a string literal in reconcilePermissionSetProjection))
• content/docs/releases/v15.mdx (via sys_permission_set (literal, a string literal in bootstrapPlatformAdmin; a string literal in reconcilePermissionSetProjection))
• content/docs/releases/v17/17-0.mdx (via sys_permission_set (literal, a string literal in bootstrapPlatformAdmin; a string literal in reconcilePermissionSetProjection))
• content/docs/releases/v17/17-1.mdx (via sys_permission_set (literal, a string literal in bootstrapPlatformAdmin; a string literal in reconcilePermissionSetProjection))
• content/docs/releases/v17/17-5.mdx (via sys_permission_set (literal, a string literal in bootstrapPlatformAdmin; a string literal in reconcilePermissionSetProjection))
• content/docs/releases/v17/17-6.mdx (via sys_permission_set (literal, a string literal in bootstrapPlatformAdmin; a string literal in reconcilePermissionSetProjection))

content/docs/releases/ is RELEASE-OWNED (AGENTS.md "Documentation Guardrails"): release
notes are written centrally at release time, and a code PR that edits them is the exact PR
that guardrail exists to stop. They are still audited — read-only. If one of them is actually
wrong, file an issue or open a dedicated docs-only PR; do not edit it here.

What this run could not see

• the SDK route bridge reached 54 of 206 client-bound route-ledger rows — the other 152 have no registrar path: tail to select them, so pages documenting THEIR client methods cannot appear above, on this or any run. Of those 152: 0 are remediable by widening that discovery convention (an in-repo file declares the path; the convention did not scan it); 55 are structural — on a ledger where NOT ONE row is declared in-repo, so no discovery change reaches them at any price; 97 are undecided (no in-repo declaration, on a ledger that has other in-repo registrars — absence and an unreadable spelling are not distinguishable here). The rows themselves: node scripts/docs-audit/affected-docs.mjs --bridge-coverage
• a page that states a rule by its inputs shares no identifier with the emitter that implements the rule, so an emitter-only diff cannot list it — not on this run and not on any run. Measured on fix(driver-sql): emit varchar(maxLength) for a text field a declared index keys on #11430: content/docs/protocol/objectql/types.mdx documents the text-family column mapping by the ObjectQL type names it maps FROM (text / textarea / html) while the diff changed createColumn; it went unlisted, and it was the page that diff falsified, in four places. No shared token exists to detect this on, so a rule your change carries has to be re-read by hand in the pages that restate it.
• a key NAME is not a key, so the hand re-read the line above prescribes can land on the wrong schema. The same spelling is authorable on one governed type and a [REMOVED] tombstone on another for each of active, aria, joins, objects, template, tools and version (censused on [finding] tools is a key on BOTH AgentSchema (tombstoned, dead) and SkillSchema (live, cloud-attested), so a name-based search attributes skill examples to the agent key — it produced a false stop-the-line alarm on PR #19059 #19093 over the liveness ledger's governed types, top-level keys); nothing in a search result distinguishes the two, so a grep hit on a LIVE example reads as evidence about the DEAD key. Measured on fix(spec): the agent.tools liveness row says dead — it claimed live on a key the schema tombstoned #19059: content/docs/ai/agents.mdx was reported as contradicting the agent.tools tombstone over its tools: example at :161, which is inside the defineSkill({ block opened at :155 — the page was already correct. Settle ownership by PARSING the value against both schemas, never by the name: that literal PASSES SkillSchema, and as an AgentSchema it FAILS at tools with the tombstone prescription. ⛔ These names are not the whole class — a key retired through a .strict() guidance map leaves no tombstone in the walked shape and none of them here (tool.category, live as AIToolDefinition.category).

Coarse fallback — 16 page(s) merely mention a changed package (the pre-#9192 predicate, kept for the deliberately-wide backstop): node scripts/docs-audit/affected-docs.mjs --json 7d7943dd0dfba6c98afa2200a08983ec85f9849a → packageMentionDocs.

Which tree this was computed on

This run read content/docs from 4919987a48ed68b73bdaed5dd6420e0e705aa666 — the merge of head e41760d89c605042380022ab8e00fa27033074d4 into base 7d7943dd0dfba6c98afa2200a08983ec85f9849a, which is what actions/checkout gives a pull_request run. Not the PR head.

A worktree cut from an older main holds a different content/docs, so re-deriving there can legitimately return a different list — that is a different tree, not a wrong row. To answer on the same tree:

# while this PR is open — GitHub drops the merge commit once it closes
git fetch origin 4919987a48ed68b73bdaed5dd6420e0e705aa666 && git checkout 4919987a48ed68b73bdaed5dd6420e0e705aa666
# afterwards, rebuild it from the two parents, which stay fetchable
git fetch origin 7d7943dd0dfba6c98afa2200a08983ec85f9849a e41760d89c605042380022ab8e00fa27033074d4 && git checkout -B drift-repro 7d7943dd0dfba6c98afa2200a08983ec85f9849a && git merge --no-ff e41760d89c605042380022ab8e00fa27033074d4

node scripts/docs-audit/affected-docs.mjs --json 7d7943dd0dfba6c98afa2200a08983ec85f9849a

⚠️ That checkout carried uncommitted changes, so the commit above does not fully identify what was read.

Advisory only, and a precision-first one (#9192): a page is listed because it names a
symbol, wire route or SDK method this diff touched — not because it mentions a changed
package. Each row says which anchor put it there, so a wrong row is reportable rather than
merely annoying. To re-verify, run the docs-accuracy-audit workflow scoped to these files:
node scripts/docs-audit/affected-docs.mjs 7d7943dd0dfba6c98afa2200a08983ec85f9849a → pass the list as
args.docs, on the commit named under Which tree this was computed on.