Bulk operations — multi-select, batch edit, exports, contact sheets

Ce contenu n’est pas encore disponible dans votre langue.

Context

Game studios manage 10k–500k+ assets. The current Artist Alley post detail flow assumes you act on one post or one asset at a time. Existing DAM tooling ships deep bulk-operation surfaces: multi-select edit, batch tag, batch delete, multi-row metadata import via CSV, configurable CSV export of search results, and printable contact-sheet generation. The audit (2026-05-30) flagged these as the second-highest impact gap — anyone with > ~1k assets discovers it inside a week.

Decision

Add Phase 1.27 — Bulk operations — as a horizontal surface that hangs off the browse feed and post views, not as a separate page. The cost of a feature like this is dominated by getting the UI patterns right and the safety rails strong (you do NOT want a “bulk delete 50k assets” button without a strong confirmation).

Selection model

Browse feed grows a checkbox in each card / row. Headers gain “select all on page” + “select all in current filter” + “clear selection.”
Selection state persists across pagination within a session.
Selection is visible at all times via a floating action bar that shows the count and the available actions, like Gmail.
A “saved selection” can be stored on a collection so a curated batch can be re-acted-on later.

Available bulk actions

Tag / untag — apply a tag to N posts; chip-style entry, dry-run preview shows the diff before applying.
Set metadata field — choose a field, set a value; respects per- field validation. Mass overwrite is irreversible without versioning (see ADR 0017 tier behaviour + future version history).
Move to collection — bulk-add to one or more collections.
Change workflow state — bulk-transition (e.g., draft → approved) subject to per-user capability.
Delete — moves to trash with a configurable retention (default 30 days). Hard delete requires a second confirmation; admins only.
Download as zip — server-side packs the selection, streams as a single archive. License-tier gates concurrency.
Export CSV — operator picks columns; CSV is computed via the search engine, not the post table, so the same column shape works on a saved-search result set.
Generate contact sheet — configurable grid (rows × cols), per- thumbnail metadata footer, header / footer text, page orientation, output PDF.
Apply share link — mint a single share link covering all selected items (Phase 1.26 wiring).

Safety rails

Every bulk action shows a preview: “this will affect 4,123 items.”
Destructive actions (delete, overwrite-metadata) require typing the count to confirm: “type 4123 to confirm.”
All bulk operations are submitted as a single job-queue job. The job is paused / cancellable / resumable. Progress streams to the UI.
Audit log entries record the bulk operation as one row with the selection IDs preserved, so a single “undo” can revert the batch (where the action is reversible).

Consequences

Positive

Removes the largest single source of “this is unusable at scale” friction.
Reuses existing primitives: search engine for selection, job queue for execution, audit log for traceability.
Contact sheet generation is a single converter; ImageMagick + the existing preview pipeline cover it without a new worker.

Negative

“Undo” semantics on irreversible operations (hard delete) are necessarily limited. Default-trash mitigates the common case but doesn’t help “I overwrote 4,123 description fields.” Mitigation: a light per-field history (touched in ADR 0017’s tangled-derivation consumer list as fieldHistoryCap) so bulk overwrites are rollback-able for Pro+.

Amendment 2026-09-04 — the four-mode batch metadata edit, as built

Sprint 20c builds the Set metadata field action above, and only it, for assets. Everything else in the Available bulk actions list is untouched. What shipped diverges from the 2026-05-30 design in ways that are decisions rather than omissions, so this amendment records them against the code (#1173, #1119).

The four modes, and why the locked-fields cascade does not transfer

overwrite, fill_empties, append and remove, over ONE field with ONE proposed value.

append and remove are multi_select ONLY. The other ten types — text, longtext, rich_text, select, tree, number, boolean, date, datetime, reference — refuse them batch-wide with 422 mode_not_supported_for_type rather than being given an invented set semantics. “Append to a number” has no meaning two people would agree on, and inventing one so that four mode names look uniform across eleven types is how a catalogue ends up full of concatenated dates.

There is no batch CLEAR. An empty overwrite and a clear are different operations: an empty overwrite is a SET (the row exists afterwards, its set_at advances, a history row is written, and R1’s requiredSetRefusal governs it), while a clear REMOVES the row and requiredClearRefusal governs it. The one exception is remove emptying an OPTIONAL multi_select, which deletes the row because writing [] into value_options is a shape the single-target writer refuses.

The Decision’s “respects per-field validation” is honoured by obtaining the shipped rules rather than restating them, so the locked-fields cascade the original design implies never arises: there is no second copy of read_only, the pattern, the required rule, the vocabulary rule or the mirrored-column rule to keep in step.

SYNCHRONOUS AND BOUNDED — superseding the job-queue choice, for these four modes only

The Decision says “all bulk operations are submitted as a single job-queue job… progress streams to the UI.” For these four modes that is superseded. They execute synchronously inside the request, and:

THERE IS NO LIVE PROGRESS. The apply response is completion and result reporting.
The ceilings are the boundary. At most 500 selection entries, checked before any membership query runs, and at most 1,000 distinct expanded targets, which is authoritative. Both refuse with 422 (selection_entry_ceiling / expanded_target_ceiling) rather than trimming — a partial expansion would silently write a different set than the operator selected — and a single post whose membership alone exceeds the expanded ceiling is refused on the same terms.
The over-ceiling refusal names the TRUE distinct expanded count, not the bound it was measured against. The count is computed in the database and the target ids are read only once it is known to fit, so the server neither executes a partial batch nor pulls an unbounded id set into memory to find out how far over the selection reaches. An operator told “1,001” when the real answer is 50,000 would trim one post at a time towards a target that was never within reach.
#39’s remaining actions stay job-oriented. Delete, zip, CSV export and contact sheets are unbounded by nature; this supersession does not reach them.

Measured at the 1,000-target ceiling: p95 apply 0.83 s against a 10 s budget, one rebuild_asset_search_text and one pg_notify per written row, four post-document rebuilds rather than one per written row, and a 39.6 KB audit envelope against a 128 KB budget. A concurrent ordinary single-target write to another field ON A BATCH TARGET now queues for the batch’s duration rather than passing through it. See the 2026-09-06 amendment: the earlier figures on this line, and the framing that made “an unrelated field” the relevant variable, were both wrong.

The public wire contract

Collections are outside this plane, on both sides. They are not a selection kind, and a field whose subject_kind is collection is NOT FOUND on this endpoint — refused before any stored target value is inspected, with no token issued. The single-target asset writer does not ask this question (it writes asset_field_value whatever the definition says, while its collection twin gates the discriminator explicitly); that asymmetry is pre-existing, is NOT changed here, and is not a reason to carry the gap onto a plane that reaches a thousand records at once.

Preview takes the mode, the field, the proposed value and a TYPED selection of {kind, id} entries where kind is asset or post. The SERVER expands posts through membership, dedupes to distinct assets, and orders BY ASSET ID — not by selection order (client-supplied) and not by sort_order (mutable), because asset id is the only ordering both endpoints derive identically. Collections are not a selection kind. Preview returns the six partition counts, a per-target partition with a machine reason where one applies, a per-would_change-target concurrency token (that value row’s own set_at), an echo of the mode and the resolved canonical value, and an opaque token with an expiry.

Apply takes EXACTLY THREE MEMBERS: the token, a required operator reason, and a confirm_count that is required for overwrite and remove and FORBIDDEN otherwise. It sends no mode and no value — both are read from the token — which is why no payload-mismatch refusal code exists. One must not be reintroduced, and the value must not be added to apply in order to keep one alive.

The operator reason, and its TWO DISTINCT BOUNDS

REQUIRED, where both shipped operator-reason fields elsewhere in the API are optional. The divergence is deliberate: a bulk mutation reaches up to 1,000 records and there is no undo.

	semantic limit	raw defensive ceiling
measured on	the value AFTER trimming	the value AS RECEIVED
unit	Unicode code points	Unicode code points
value	500	2,000
on violation	400 `reason_too_long`	400 `reason_payload_too_large`
in OpenAPI	stated in the description as authoritative	encoded as `maxLength`

The schema’s maxLength is the RAW ceiling and never the semantic one, so a conforming external validator never rejects a value the server would accept — whitespace around a 500-code-point body is valid. Handler order is: raw ceiling, trim, requiredness, semantic limit. Code points, not bytes. Recorded verbatim after trimming.

⚠️ maxLength is enforced NOWHERE in generated code. Every bound is enforced in the handler or it is not enforced at all.

The token lifecycle

Bound to its CALLER. SINGLE-USE. It binds the ordered target set with each target’s partition, the field’s identity and configuration fingerprint including the vocabulary options document, the mode, and the exact canonical value. It is never authority — not to write, not to mint a term, not evidence that a reference target is still alive; all of those are re-evaluated against the current world at apply.

Token consumption, the durable field and vocabulary mutations, and the operation’s single audit envelope are ONE ATOMIC COMMITTED OUTCOME.

There are exactly two durable results and no third. A pre-write refusal commits no field value, no term and no envelope, and LEAVES THE TOKEN USABLE — so an operator who mistypes the count or omits the reason corrects and retries without re-previewing. A committed apply — including a partial one, and including one where would_change was zero — commits its result, exactly one envelope and the consumption together.

A lost HTTP response therefore never makes a spent token spendable, and a 200 is not the consumption boundary. A would_change == 0 apply is a REAL operation: it completes, consumes its token, and writes one envelope recording the zero-change operation with its reason and counts.

The complete validation precedence

No token-bound fact — the mode, the would_change count, the field, the target set, the expiry, the consumption state or the expected confirmation count — may influence any externally visible response until integrity AND caller binding have BOTH succeeded.

Token-independent request-shape checks may run first, because they leak nothing. Then, in exactly this order: integrity → caller binding → consumed → expiry → mode-specific confirmation → current authority, definition, configuration, vocabulary and reference liveness → the committed apply.

Malformed, unknown, tampered and another caller’s tokens collapse to ONE BYTE-IDENTICAL 403 preview_token_invalid, and stay indistinguishable however else they differ. The attack this closes is specific: apply does not send the mode, so a server that validated confirm_count against the token’s mode before checking ownership would answer confirm_count_required versus confirm_count_not_applicable about SOMEBODY ELSE’S preview. Repeat for expiry and consumption and it is an enumeration oracle over every preview on the instance, built entirely out of refusals.

CONSUMED WINS OVER EXPIRED for a caller-owned token, because it tells the operator their operation ALREADY RAN where preview_expired would invite a duplicate run.

The status taxonomy

400 — the request could not have been valid for ANY state of the system.
403 — authorization, INCLUDING anything not provably this caller’s token.
422 — well formed and understood, and this definition, value or scale refuses it.
409 (apply only) — the token is provably this caller’s own but is spent or stale, or the world moved; the remedy is to re-preview.

dangling_reference (preview: it never resolved) and reference_invalidated (apply: it resolved when the operator looked and has since stopped) are deliberately NOT collapsed.

The counts, and what the typed confirmation confirms

expanded = would_change + no_op + refused + inapplicable + unreadable + unauthorized
eligible = would_change + no_op
would_change = changed + conflict + gone + unauthorized_at_apply + error   [at apply]

Every target of a successful preview belongs to EXACTLY ONE partition. The Decision’s “type 4123 to confirm” is honoured, and the number is would_change, not eligible: the number an operator types is the number of records that will actually change.

overwrite reports no_op as ZERO even against a target already holding the proposed value — a set advances set_at and writes a history row, so it changes the record even where it does not change the value.

The five-gate stack, and read-as-well-as-write

	gate	resolution
G1	the bulk instrument in the target’s scope	per target
G2	the ordinary subject authority rule	per target
G3	applicability	per target
G4	the field’s own `write_capability`	batch-wide
G5	the field’s `read_capability`	per target

Only below all five is any stored value inspected. G5 is on the WRITE path because the operation reports: a batch that said “twelve of your targets already hold this value” would have answered twelve questions about a field the caller may not read. An unreadable target discloses nothing — not emptiness, not membership, not equality, not set_at, and not in the audit envelope.

Neither the bulk capability nor the subject rule replaces the field’s own write gate, and it does not replace them.

Apply re-checks G1, G2 and G5 per target and G3, G4 batch-wide, against EFFECTIVE permission and never raw grant-set equality — a caller who loses one direct grant while a role still confers the capability has lost nothing. A caller whose GLOBAL bulk grant is revoked while a SCOPED grant still covers part of the selection is NOT failed wholesale: the covered targets proceed and the rest become unauthorized_at_apply.

assets.metadata.bulk_edit is TEAM-SCOPE-AWARE, and a TEAM-LESS asset requires a GLOBAL holding — the nullable trap visibility.MayMutate documents. The field’s own write_capability and fields.vocabulary.extend are reproduced at their SHIPPED GLOBAL-ONLY scope; that asymmetry with the team-scope-aware read gate is real, is NOT fixed here, and is recorded by a standalone preservation test so a later fix moves both together.

The four atomicity families

“Same transaction” is NOT sufficient at READ COMMITTED. The precondition lives on assets and the mutation lands on asset_field_value, a DIFFERENT TABLE, so 20a’s single-statement guarded-write pattern does not transfer. The foreign key’s implicit FOR KEY SHARE protects none of it — it conflicts only with FOR UPDATE, while ownership transfer, team move and soft delete all take FOR NO KEY UPDATE — and there is no foreign key on value_ref at all.

Per-target subject. The target’s current owner, team and liveness, the caller’s G1/G2/G5 verdicts, and the mutation they authorise are one atomic operation relative to competing transfer, team move and soft delete. Failure is per target.
Batch-wide definition, configuration and vocabulary. Exactly two valid serial outcomes: the external change wins and the batch refuses with ZERO writes and no mint, or the batch wins and the ENTIRE batch executes under one validated state. Never the first N under the old rules and the rest under the new.
Proposed-reference liveness. A pre-batch re-check is not sufficient; it establishes a fact that can stop being true before the last write lands.
Effective authority. The caller’s effective verdict and every mutation it authorises are one atomic operation relative to a competing authority change. This covers ALL FOUR authorities the batch consumes — bulk-edit admission and per-target scope, subject authority, the field’s own write capability, and mint authority — because all four are drawn from ONE effective-authority read.

Technique is implementation-owned. As built: ONE ascending FOR UPDATE statement over the would-change subjects UNION any proposed reference target, taken before a single value is written; FOR UPDATE on the field definition taken BEFORE it is read; and, for authority, a transaction-scoped ADVISORY LOCK taken BEFORE the authority read and held to commit. The subject and reference locks were FOR SHARE, taken one target at a time; the 2026-09-06 amendment records why that deadlocked and what replaced it.

Amendment 2026-09-04 — re-resolving authority in the transaction is NOT serialization

The first implementation of family 4 re-read the caller’s capabilities from the apply’s own transaction and treated that as sufficient. It is not, and this ADR said so about every other family while missing it here.

At READ COMMITTED each statement takes a fresh snapshot. Re-reading in the transaction makes an authority change committed BEFORE the read visible, and does nothing about one that commits AFTER it: that change lands while the verdict is still being relied upon, and the writes it authorised go through. The batch locked the field definition, the reference target and each subject — and authority lives in none of those. It lives in user_roles, roles, role_capabilities, user_capability_grants, user_capability_revokes and team_closure.

⛔ A row lock could not have closed it either. The dangerous mutation is frequently an INSERT — a revoke ADDS a row to user_capability_revokes — and FOR SHARE locks rows that exist. PostgreSQL has no predicate locking at READ COMMITTED, so no locking of the rows the authority query reads can exclude a row that does not exist yet. An advisory lock is not a shortcut here; it is the only mechanism that can express the claim.

As corrected. One transaction-scoped advisory lock space, with a reader half and a writer half. The batch takes the SHARED half — on the caller’s key and on a structural key — BEFORE resolving authority, and holds both to commit.

⛔ Both sides participate, and that is the whole mechanism. A lock only the batch took would prove nothing — which was the defect in the original evidence for this family, whose test made its competing revoke wait on a field_definition lock that no production revoke touches.

The authority-writer inventory, stated as it is

An earlier draft of this amendment said “every production path that mutates authority takes the exclusive half”. ⛔ That was a universal claim reached before the enumeration was complete, and it was false: requests.Handler.Grant writes user_capability_grants from the requests package and had been missed. The inventory is therefore given in full, with each entry either PARTICIPATING or carrying a stated lifecycle exemption.

Participating — take the exclusive half before the write:

writer	scope	key
`auth` · add a per-user grant	transaction	user ref
`auth` · remove a per-user grant	transaction	user ref
`auth` · add a per-user revoke	transaction	user ref
`auth` · remove a per-user revoke	transaction	user ref
`auth` · assign a global role	transaction	user ref
`auth` · capability expiry sweeper	transaction	structural
`requests` · approve a capability request	transaction	requester’s ref, not the approver’s
`teams` · add a team parent	transaction	structural
`teams` · remove a team parent	transaction	structural
`cmd/aa` · `aa seed` bootstrap admin	session	structural
`cmd/aa` · `aa seed --reset` (`resetContent`)	session	structural
`seed` · `Runner.applyTeams` (`team_closure`)	session	structural
`seed` · `Runner.applyFixturePrincipals` (`user_roles`)	session	structural

The sweeper, the team-parent paths and every seed span use the structural key because their blast radius is not one nameable user: the sweep reaps across every user at once, re-parenting changes what every team-scoped grant expands to WITHOUT TOUCHING A SINGLE GRANT ROW, and the reset empties the authority tables outright.

`aa seed` is NOT lifecycle-exempt, and the session scope is why

An earlier draft of this inventory wrote it off as “offline maintenance that also TRUNCATEs”. ⛔ Both halves were false, and the mistake was the same one that produced the false universal above: classifying by command name instead of by call site and actual concurrency.

It is DESIGNED for a live instance. Its migrate step is documented as safe “whether the server already migrated, is migrating right now, or was never started”, and --reset broadcasts a wildcard cache flush over NOTIFY precisely because “the seeder is a separate process with no cache Registry” and a server may be serving throughout. And the TRUNCATE is the BLAST RADIUS rather than a reason to dismiss it: seed.Reset’s TRUNCATE ... CASCADE and its DELETE FROM teams empty user_roles, user_capability_grants and user_capability_revokes wholesale, and bootstrap.Run then restores the admin’s role.

⛔ The seed spans take the lock at SESSION scope, and that is not a detail. Their authority mutations run as a series of autocommit statements — a TRUNCATE, several DELETEs, a bootstrap restoration, then the runner’s own writes. A transaction-scoped lock would be released at the end of whichever statement took it and would leave the rest of the span unprotected: a lock that exists, that production really takes, and that does not span the thing it protects. Session and transaction advisory locks share one space and conflict identically; scope decides only when a lock is released.

⛔ A session lock makes its own release a correctness problem, not a tidiness one. pgxpool.Conn.Release() returns a connection TO THE POOL FOR REUSE, and a session-scoped advisory lock survives until it is explicitly unlocked or the PostgreSQL session is destroyed — an ordinary release does neither. So a single failed unlock (a cancelled context, a network blip, a statement timeout) would hand a still-locked session back to the pool, where it would sit idle and unowned holding the structural lock forever, and every batch apply after it would block indefinitely on a lock with no owner. A swallowed error there is a total, permanent stall of the feature with nothing to explain it.

The release therefore HIJACKS the connection out of the pool and CLOSES it whenever the unlock fails, which destroys the backend session and is what actually drops the lock, and logs at ERROR. The pool opens a replacement on demand.

Four separate, non-overlapping spans, each no wider than the authority mutation it protects. An earlier version wrapped the seed runner’s whole Run, which held the STRUCTURAL lock — the one that excludes every batch authority reader — across catalogue loading, users, memberships, follows, fields, collections, featured, ASSETS, POSTS, likes and comments. A reseed runs for minutes, so an unrelated batch apply could wait out its entire deadline while image files were loading. Only two of those phases mutate authority.

The runner’s protection lives IN the two phases rather than around the call, so any caller gets it — including a test that drives one directly, which is how the narrowing is proven rather than asserted.

⭐ The phase, not the statement, is the semantic unit for both, and that is a claim rather than a convenience. team_closure describes a HIERARCHY: it is only the shape the catalogue specifies once every team’s rows are in, so a reader resolving a scoped grant against a half-built closure sees a real and different expansion. The fixture principals are seeded as a SET, and a reader midway through sees a world that is neither the old one nor the new one. Locking per statement would open exactly those windows.

The exhaustive check. Every write to user_roles, roles, role_capabilities, user_capability_grants, user_capability_revokes, team_closure and team_parents in the seed package resolves to those two queries and no others. team_memberships is deliberately absent: EffectiveScopedCapabilitiesForUser does not read it, so it is not authority-bearing for this invariant.

⚠️ The operational consequence, stated rather than absorbed. While the structural lock is held, batch metadata applies WAIT — so what matters is how long it is held, and it is held only for the authority spans above, never across a whole command. The long parts of a reseed — catalogue loading, users, memberships, follows, fields, collections, featured, ASSETS, POSTS, likes and comments — run WITHOUT it.

An earlier revision wrapped the entire Runner.Run, and did have the consequence an earlier version of this paragraph described: an unrelated apply could wait out its whole request deadline while image files were loading. After the narrowing that is no longer true, and the wording is corrected rather than left standing — a stale operational warning invites a fix for a problem that is not there.

The remaining waits are bounded by the reset and bootstrap spans, and a collision fails CLOSED with nothing written. A lock_timeout on the batch’s acquisition is deliberately NOT taken: it would change the batch’s failure taxonomy, and the excessive-scope problem it would have papered over is fixed at the source instead.

Exempt, by construction — an in-flight batch for the affected principal cannot coexist with these:

writer	why coexistence is impossible
first-boot setup assigns the initial admin role	the handler CREATES the user in the same request; a principal that does not exist cannot have a batch in flight
self-registration assigns the default role	same — the row is inserted immediately above, and the account cannot authenticate until it is verified
server-startup `bootstrap.Run` (`run()`)	executes after migrations and BEFORE the HTTP server accepts anything, so no request of any kind is in flight

⭐ One function, two call sites, two answers. bootstrap.Run appears three times. The startup call is exempt for the reason above. The two inside aa seed are a different concurrency context entirely and are covered by the seed spans in the participating table. The exemption belongs to the CALL SITE, never to the function.

⭐ These are exemptions with reasons, not omissions. If any of them ever gains a path that can run against a live principal, it joins the table above.

Effective-capability semantics are unchanged: role-derived authority, direct grants, revokes, team-scoped grants and closure inheritance all still decide the verdict, and the decision is never reduced to raw grant-row equality — a team-scoped ROLE assignment produces zero rows in user_capability_grants, which is why ScopedTeams exists.

Vocabulary

A PREVIEW MUST NOT MUTATE. The shipped openOrCheckVocabulary mutates by design, so the preview resolves through the same index and resolveOrMint with creation removed, and REPORTS mintable terms rather than creating them.

The verdict SPLITS. Canonicalisation, unknown-slug handling and mintability are BATCH-WIDE. The deprecated-or-archived grandfather test needs the target’s own held set — the same slug is grandfathered on a target holding it and refused on a sibling that does not — so it is per target and produces a target-level refused.

THE MINT IS COUPLED TO A REAL WRITE. A new term commits only if at least one successful batch mutation actually stores it. “Preview predicted would_change > 0” is insufficient: if every such target ends conflict, gone, unauthorized or error, the options document stays BYTE-IDENTICAL. The cache is invalidated once, after commit, only when a term committed.

Preview binds the EXACT canonical slug set apply stores, which is what prevents a PHANTOM would_change: a term that looks new but canonicalises onto a slug the target already holds counts no_op at preview and stays one at apply.

Per-type semantic emptiness

Obtained from the shipped valueIsEmpty, all eleven types. FALSE is a real boolean value and is never empty. rich_text is sanitised FIRST and judged on the output, because the sanitiser strips no empty elements.

select and tree DO NOT treat whitespace like text, and the split is the shipped behaviour rather than a batch invention: "" returns nil from vocabularySlugs and never enters the vocabulary pipeline, so it stores as '', while " " enters it as a slug, matches nothing, and is refused as unknown. Nothing trims: a text field given " " stores " ".

fill_empties with a semantically empty proposed value is refused BATCH-WIDE on required AND optional fields, across every type — the mode means “give the empty ones a value”, and a value that is itself empty makes it a contradiction. No mode creates an accidental empty row.

Archive is not deletion, on BOTH planes

Only deleted_at IS NOT NULL removes a subject from the authority probe, and only deleted_at IS NOT NULL invalidates a reference target. An ARCHIVED asset is written, and an ARCHIVED asset is a valid reference target.

Audit

ONE envelope per apply, ever, committed in the same durable outcome as the writes and the consumption. It records the initiating human, the operation id tying preview to apply, the mode, the field, THE TRIMMED REASON VERBATIM, the accepted confirmation count, all six partition counts, both reconciliation equations, the outcome counts with their sub-reasons, the terms that ACTUALLY committed, and the resolved target ids.

It records NO FIELD VALUE, old or new. The audit log is read under system.audit.read, a DIFFERENT POPULATION from the field’s readers, so putting values there would be a side channel around the field’s own read gate — a thousand records at a time. Every value is already in asset_field_value_history, behind that gate. Unreadable and refused targets contribute their id and a non-value-sensitive partition label only.

Unlike the Decision’s “a single undo can revert the batch”, there is no batch undo in this subset. The per-field history the Consequences anticipate exists and is written per target; a bulk revert built on it is not part of this amendment.

Two narrow wire-hardening codes, discovered in implementation

unknown_mode and unknown_selection_kind, both 400. Neither is part of the batch’s product surface; both exist because nothing between the wire and the handler validates a closed enum. There is no spec-validation middleware, and the generated enum types are bare strings whose Valid() has no callers, so an undefined mode or selection kind arrives intact. Left unchecked, an undefined mode matched no mode-specific arm — which let a REQUIRED field take a semantically empty value without tripping R1, partitioned every target as a no-op, and then died on a CHECK constraint as a 500.

They are recorded here so the enum is not re-derived without them, and so the next reader knows they answer a shape question rather than a product one.

Deliberately out of scope

The single-target subject-authority gap (the ordinary field-value writer still has no asset-level ownership check — tracked separately), the write_capability team-scope fix, a foreign key on value_ref, a required-bypass permission, batch undo, a generic batch clear, collections as a selection kind or value plane, select-all-in-filter, and the frontend selection and batch surfaces.

Cited, not re-decided

ADR 0052, ADR 0012 (per-type emptiness and the rich-text survivor list), ADR 0099 §§1, 5 and 8 (a hidden field is still writable; the non-disclosure filter; the atomicity precedent), ADR 0083 §2, ADR 0092 §2 (fields.vocabulary.extend as a dial).

Amendment 2026-09-06: the lock hierarchy, after a real deadlock

Sprint 20c-i’s apply deadlocked against an ordinary single-target metadata write. SQLSTATE 40P01, intermittently, on post-merge Integration CI. It was not a test artifact and it was not contention: it was a LOCK-ORDER INVERSION, and the edge that closed the cycle sat two trigger levels below the statement the error named.

The chain

INSERT asset_field_value ... ON CONFLICT
+- asset_field_value_search_text  (AFTER INSERT OR UPDATE, per row)
   +- rebuild_asset_search_text(asset_id)
      +- UPDATE assets SET search_text ...   -> FOR NO KEY UPDATE on
         |                                      assets[V], held to COMMIT
         +- assets_member_post_search_text  (AFTER UPDATE OF search_text,
            |                                sensitivity, status,
            |                                processing_status, deleted_at)
            +- asset_member_post_search_text_trigger()
               FOR r IN SELECT post_id FROM post_assets
                         WHERE asset_id = NEW.id      -- NO ORDER BY
               +- rebuild_post_search_text(post_id)
                  +- UPDATE posts SET search_text ->  FOR NO KEY UPDATE
                                                      on posts[P]

EVERY metadata write therefore takes an assets row, and then a posts row for each containing post that asset has. The batch, which locked and wrote one target at a time, held a post row from an earlier target while asking for the next asset row:

batch    holds posts[P]  -> waits assets[V]
ordinary holds assets[V] -> waits posts[P]

Two transactions, the same two tiers, opposite order. Nothing either one did was individually wrong, which is what made the wait graph look unclosable and is the reason this is recorded rather than fixed quietly.

The correction

Asset tier. ONE statement locks the would-change subjects UNION any proposed reference target, FOR UPDATE, ORDER BY id, before a single value is written. It replaces both the per-target FOR SHARE probe and the separately taken reference-target lock, and the batch never asks for an assets row while holding a posts row again.

FOR UPDATE rather than FOR SHARE, and the reason is the same trigger chain that caused the deadlock. Every write ends in an UPDATE of the subject’s assets row, which needs FOR NO KEY UPDATE, so a FOR SHARE holder is a holder that must UPGRADE: two batches sharing one target would each hold FOR SHARE and each block on the upgrade. FOR UPDATE is also the weakest mode that makes a membership INSERT queue rather than race, because a foreign key takes FOR KEY SHARE and that conflicts with FOR UPDATE and not with FOR NO KEY UPDATE.

Post tier. The per-row asset-to-post propagation is SUPPRESSED inside the batch transaction, with a SET LOCAL flag the trigger checks, and the debt is paid once per DISTINCT containing post, ASCENDING, after all writes. A thousand targets across four posts is four rebuilds. The suppression is transaction-local by construction, so it cannot survive into the next caller that borrows the connection.

The shared rebuild function locks before it reads. rebuild_post_search_text takes its post FOR NO KEY UPDATE as its FIRST statement. It bakes its document out of three reads that all precede its UPDATE, and at READ COMMITTED blocking on the UPDATE’s row lock re-evaluates the target row and NOTHING ELSE: the local variables keep whatever snapshot they were computed from. Without the entry lock a transaction can compute a post document, wait, and then apply it over a document another transaction just committed. That is reachable from a membership removal and from the addition of a member that is not itself being edited, and both now have regressions.

The multi-post loops are sorted. asset_member_post_search_text_trigger, assets_mature_sync and assets_ai_provenance_sync walk their posts ORDER BY post_id. The value of that sort is INDEPENDENT of the batch editor: two ordinary writers, editing two assets that belong to the same two posts whose membership rows were written in opposite orders, deadlocked with no batch anywhere near them. Migration 00067.

The hierarchy, stated as an order

field_definition[field], FOR UPDATE, row-scoped, and the shared half of the authority advisory lock.
Subjects UNION reference target: FOR UPDATE, ascending assets.id, ONE statement.
The field-value writes, and the asset search_text rebuild they set off: NO KEY UPDATE on rows already held at tier 2, and NO POST LOCK.
The coalesced post rebuild, ascending post_id, each call taking NO KEY UPDATE on its post at entry.
Vocabulary mint: FOR UPDATE on the vocabulary row, which is the field_definition row already held at tier 1.
The audit envelope: an append-only insert whose foreign keys take KEY SHARE on rows already held.

Acyclic, because the asset tier closes before any post lock is taken, no path takes an assets row lock after a posts row lock, and every multi-post acquirer walks ascending.

Tier 3 holding no post lock is load-bearing and is not free. assets carries two other AFTER UPDATE triggers, assets_mature_sync and assets_ai_provenance_sync, which both loop post_assets and touch posts. Neither fires here only because both early-return unless mature, ai_provenance or the nullness of deleted_at actually changed, and a field-value write changes none of them. That is an assumption about somebody else’s trigger, so it is covered BEHAVIOURALLY: with a real batch parked mid-write-phase, an independent transaction takes FOR NO KEY UPDATE NOWAIT on every containing post and must get it.

What this costs, measured

At the 1,000-target four-post ceiling, on one host, 20 samples each:

	before	after
apply p95, uninstrumented	20.2 s, of which 18 runs sat at 3.5 s and two at 20 s and 22 s	0.83 s, every run within 35 ms of the median
apply p95, under `-race`	3.8 s	1.19 s
ordinary write to another field ON A BATCH TARGET, during the batch	7 ms	19 ms in one run, 829 ms in another

The apply got several times faster and the spread collapsed, which follows from coalescing: one 250-member post-document rebuild PER WRITTEN ROW becomes four for the whole operation. The pre-correction run’s two outliers at 20 s and 22 s are recorded as measured; their cause was not established.

The contention figure moved the other way, and it is now a coin flip rather than a constant. That is the intended trade and the honest description of it: the batch holds every target FOR UPDATE for the length of the operation, so a concurrent write to a TARGET either starts before the batch reaches its lock pass and passes through, or arrives after and waits.

A write to an asset the batch is NOT editing takes none of the batch’s asset-tier row locks, so it never contends at tier 2. It can still wait on the batch later: if that asset shares a containing post with any batch target, its own trigger chain requests that post while the batch’s rebuild phase holds it. Direct tier-2 contention is confined to batch TARGETS; contention mediated by a shared post is not. The transaction-wide regression exercises exactly that path, with an ordinary write to a non-target that shares a post with one.

The ceiling test only LOGS this number; it asserts that the concurrent write does not error and that its value lands, and that contract is unchanged.

The old line in this ADR, “a concurrent ordinary single-target write to an unrelated field completing in 4 ms”, was doubly wrong. The figure is not reproducible, and “unrelated field” was never the relevant variable: the trigger chain runs identically whatever field was written, so the contention was always about the SUBJECT and its posts and never about the field.

Preserved unchanged

The validation precedence and the zero-change rules above are untouched by this work and are not re-decided here. A would_change == 0 apply is still a REAL operation: it still runs the field-definition seam and the authority serialization, still consumes its token, and still writes exactly one envelope. What N = 0 removes is only the work there is none of, namely subject rows to lock and post documents to rebuild, and it is NOT a bypass: a proposed reference target still joins the asset tier’s lock set, so a reference that has gone still refuses batch-wide.

Out of scope, and recorded rather than fixed

recompute_post_mature and recompute_post_ai_provenance share the read-before-lock shape corrected in rebuild_post_search_text. The batch never changes their inputs, so there is no freshness interaction with this work and they are left alone. They are named here so the next reader does not have to rediscover them.

Alternatives considered

CLI-only bulk surface. Doesn’t help the reviewer / coordinator audience. Rejected.
Reuse search export only, no UI selection. Doesn’t solve bulk edit, only bulk export. Rejected — incomplete.
Per-action separate pages. The older DAM pattern; the UX is dated. The modern pattern is one floating action bar over the existing browse view. Adopted.

Reference

Phase 1.27 in docs/roadmap.md.
Search engine (Phase 1.12 Search 2.0) supplies the selection-by- filter mechanism.
Job queue (Phase 1.18.A — shipped) executes bulk actions.
License tier governs bulk concurrency + selection ceiling via the tangled-derivation model in ADR 0017.