Unifies Salvador vs RMS subgroup logic for territory rollups, aggregate class slugs, and the intra-TI capture benchmark so the MAUP-sensitive rule cannot drift across call sites.
Fold duplicated campanha access, loaders, list/filter/form shells, and scripts into shared modules; reorganize utilities by domain; stabilize e2e; and document prevention-first audit workflow so the same smells cannot creep back in.
Bookmarks the canonical list recorte beside Limpar, syncs via useSyncExternalStore, and matches active URLs with a pure href comparator so the app layout never imports the heavy URL serializer.
Wire CampaignListPendingBoundary/Results on /campanha/atividades/giros so TI navigation matches list pending UX; pin contract in tourRegionPending unit spec.
Research only — no roadmap items, no verticals, no code; validation with the real person gates everything (notebook: .cursor/rules/projects/emendas-mandato.mdc).
After B8 F2 the choropleth already opens each Salvador ZE; drop the textual list and zoneBreakdown from the bundle, keeping the mesh provenance caveat via hasZoneMunicipalities.
Staff sees a Flag + tooltip on the name column and mobile card instead of the destructive “Prioritária” chip, with copy centralized in municipalityLabels.
No product items registered yet — the 2026-07-29 session with the real gestora decides whether the finance track opens; pre-validation P0-P2 assessment lives in the project notebook.
Nothing left from the reviewers scored high enough to earn a roadmap item, so
the four discards are written down with their reasons instead — a discard whose
reason is missing comes back next session dressed as an obvious win. The one
worth keeping is the enrollment toast's third spelling of the failure copy: it
is deliberate, because that toast has no retry affordance and ending the
sentence in "Tente novamente." would be a false instruction.
Three engineering deferrals gained the trigger they lacked (shared HMAC core,
an optional `bodySchema` on the route shell, promoting `postJson`) — all three
are second-call-site abstractions, so the trigger is the third.
Also corrects two statements this delivery falsified: the notebook still said
`postJson` translates `SyntaxError`, which the simplify pass deleted, and the
roadmap's B40 item carried the pre-optimization 130 kB that its own header line
had already restated as 126 kB.
`postJson` caught a non-JSON response and threw a plain `Error` carrying
"Resposta inválida do servidor.", but all three islands read `message` only off
a `CampaignWebAuthnError` — so the sentence was written for the person and then
thrown down the channel that discards messages. Deleted rather than re-typed:
the islands' own fallback is better copy for a field worker, which is the same
reason the five list-cell controls substitute their own. Nothing user-visible
changes, because the string was already unreachable; the comment now records
why the branch is absent, so it does not come back.
The login failure sentence was also spelled twice for one slot on one screen —
the route's `genericMessage` and the button's fallback — so it became
`CAMPAIGN_BIOMETRIC_LOGIN_ERROR_MESSAGE`, twin of the enroll constant the
register half already shares between its route and its card.
Also widened two doc comments this branch made stale: `campaignJsonMutationRoute`
and `campaignJsonRequest` both claimed to serve only the list cells' quick edits,
and four of nine callers are now WebAuthn ceremonies.
The rebase brought a convention guard that refuses a route.ts exporting POST
outside campaignJsonMutationRoute, and it named all four WebAuthn routes. They
now come from the shell, so the same-origin check is structural rather than
copied, and postJson delegates to postCampaignJson — keeping only what the
ceremonies add on top: the envelope's discriminant decides instead of the
status, and a non-JSON response reads as one sentence while a network failure
still propagates untouched. The two *-options routes have no body, hence
campaignWebAuthnNoBodySchema.
WebAuthn platform passkeys: opt-in enrollment after a password login, a
password-less sign-in that mints the same 14-day campaign-token, and
revocation from the profile. The session is minted through Payload's public
pieces (checkLoginPermission, addSessionToUser, getFieldsToSign, jwtSign)
inside a transaction with an advisory lock, so a passkey login neither
bypasses a lockout nor evicts the account's other sessions. The rpID is
derived from NEXT_PUBLIC_SITE_URL and resolves to null when the request host
disagrees, so previews never offer a ceremony the browser would refuse.
The /simplify triage produced two live debts, and both are about the same pair
of functions the B34+ plan already names in F3, so they landed as F10 of that
plan rather than as an escala-dry-pos-b37: the município-side batch actions
write one município at a time (findByID + update each — ~870 round trips at the
435 ceiling B37 itself raised, but ~52 at the largest real TI, and no test above
a batch of 2), and the two are twins at 2 call sites, one short of the bar. The
common trigger is a "select all"/multi-TI affordance or a third batch on that
side, and F10 records why the bulk-SQL rewrite is not obviously the answer: it
would hand back the scope check that overrideAccess: false gives for free.
This delivery also closed B34+ phases outright, which is worth reading in the
plan rather than inferring from a diff. F2 is done and its prediction was wrong
in a way worth keeping: it stated "o B37 não retira este débito", and B37 read
the same evidence the other way — the ~165 measured lines were not deduplicated
into useChipRowOverflow, they were deleted when both cells became wrappers of
one engine. F9 is done via scopedPortfolioIndex. Half of F3 is done (the three
single-id deltas), and the half that stayed out — the set-with-floor form —
stayed out because seeing the collapsed code confirmed the plan's own inverted
recommendation. F5 and two F6 bullets shrank for the same reason: "written twice
across the two surfaces" stopped having two surfaces.
Also corrects a comment I got wrong in the previous commit: the batch's sort
does not carry the deadlock-avoidance guarantee, since acquireTextAdvisoryLocks
sorts its own keys (C8 batched them into one round trip); it only makes the
apply order stable.
nextStateDeputyIdsAfterMunicipalityMembership was the third byte-for-byte copy
of the same single-id delta — nextAdvisorIdsAfterMembership and
nextStateDeputyIdsAfterMembership differ from it only in which cap they read —
so the algorithm now lives once in nextIdsAfterMembership with the cap as data,
and the three named wrappers just bind their own. The three cloned spec files
became one describe.each suite: coverage went up rather than down, since every
relation now gets the "can go to zero" and no-mutation cases only B37's clone
had, plus an explicit assertion that each wrapper refuses at ITS cap with ITS
message.
That pairing matters because mapCampaignFormActionError matches thrown messages
against safeMessages by exact string, and two of those cap messages were bare
literals in three src places each (thrown, plus two independent allowlists):
rewording any one would have silently collapsed a real refusal into "tente
novamente". Both are constants now, next to their caps, following the
MUNICIPALITY_STATE_DEPUTIES_CAP_MESSAGE precedent already in those same files.
The dobradinha staff message had the same exposure in four places — and
actions/stateDeputy.ts already imported the constant while restating it as a
literal one line below — so it is STATE_DEPUTY_STAFF_MESSAGE (renamed: it
guards create/update too, not just the batch), and the conflict message it
shares with the create route is named too.
The batch action stopped resolving the dobradinha slug before its loop: the
slug exists only to target the revalidate, which the caller already skips when
nothing changed, so on a no-op batch that read was fetched and thrown away —
the guard setLeadershipMunicipalitiesMembershipRecord already had. Its
redundant second dedup is gone (the schema's transform already deduped; the
sort stays, for deadlock-free lock ordering), and the bare 435 shared with the
advisor batch is MAX_MUNICIPALITIES_PER_BATCH.
Two memoization gaps closed, both restoring invariants the codebase had already
established elsewhere. An advisor's scoped .filter() returned a fresh array per
row, defeating the WeakMap that exists precisely to share derivations across
rows, so scopedPortfolioIndex memoizes on (index, addableIds) and every row
gets one array. And LeadershipStateDeputyRelationCell called matchesAtWordStart
per candidate, paying an NFD pass plus three Unicode regexes on both sides per
keystroke over the unpaginated leadership catalog — the cost wordStartFilter's
own doc comment warns about; it now normalizes the query once against labels
cached on options, like municipalityPortfolio.
Rejected with reason: maxItems on the dobradinhas cell. A dobradinha has no
product ceiling on municípios (that is why the bound is the catalog), so a
"generous 30" would invent a rule that refuses a legitimate statewide
dobradinha, and the honest 435 is unreachable when ids are unique.
Also dropped the dead addPlaceholder knob no caller ever set.
The "Municípios" column of /campanha/dobradinhas reuses
MunicipalityPortfolioCell unchanged, writing the inverted side of the
relation (municipality.stateDeputies) through a single transactional
batch action (setStateDeputyMunicipalitiesBatch): per-município advisory
lock, all-or-nothing on a partial-scope batch instead of a silent partial
apply, no floor (a dobradinha can hold zero municípios). stateDeputiesArraySchema's
cap moved 20 → 435 to fit a whole território batch from the município side.
The post-implementation critique found two P1s, and both were fixed instead
of deferred. First: MunicipalityPortfolioCell and LeadershipStateDeputyRelationCell
had grown the same chip-editing machine twice — optimistic delta, undo, ARIA
1.2 combobox, floor/ceiling, pointer-fine/coarse split — with incompatible
contracts. This was the 3rd call site B34+ was waiting for, so RelationChipCell
is now the single engine and both cells shrank to thin domain wrappers (chip/
search building, FormData, copy — nothing interaction-specific left in either).
LeadershipStateDeputyRelationCell also gained an ownerName prop and its first
dedicated test suite, which it shipped without before this.
Second: an advisor could see, and attempt to remove, a município chip outside
the portfolio they administer, failing silently at the document level.
MunicipalityPortfolioCell now filters the visible chips by addableIds instead
of only gating what can be added, so an advisor sees and can only touch
municípios they administer.
No migration, collection, or Consent. Docs updated: plan as-built, roadmap,
AGENTS.md, project notebook.
Two findings outlived the simplify pass, both structural. Hiding a column does
not switch off the query that exists only for it: the two relation catalogs are
loaded inside the `Promise.all` and the cookie is read after it, and since
`cookies()` does no I/O, the fix is statement order rather than a new await.
And the "Colunas" button disappears with the table on the three surfaces that
swap the whole table for an empty block instead of using `CampaignTable`'s
`empty` prop, so filtering to zero rows removes a control the filter never
touched.
Five findings were discarded with the measurement or the reason beside them, and
three deferred with a trigger — the shared Popover chrome, the e2e staff fixture
and a shared "load unless hidden" helper all sit at two call sites.
The four *SortableHead islands already fall back to their sort-label record,
so B17's new `label` made the header text a third copy instead of removing the
second one — 16 redundant `children` deleted, leaving only `frescor`, which
diverges on purpose and now reads the record too.
The 3 s "safety net" timer was committing like a close, refresh included, which
recreated the per-toggle route render the 400 ms debounce was rejected for: at a
reading pace every checkbox landed outside the window. The timer now writes the
cookie only; closing repaints. A session that ends where it started refreshes
nothing, and a menu open when the route unmounts saves from the cleanup, since
Radix reports no close event for a sidebar navigation. Both pinned red first.
Also: `columnVisibility` required rather than optional for the tests' benefit,
the redundant per-list cap dropped, `resolveVisibleColumns` no longer copies,
and B29's unused `LeadershipListAccessFilter` export un-exported — knip was
failing on main too.
The rebase landed B29's sort/filter heads on /campanha/liderancas next to B17's
now-required `label`, and the two do not substitute for each other: B17 dropped
heads that only restated a label, but these carry the sort and filter controls,
so both sides survive on the same column. `updatedAt` is new from B29 and had
no label at all.
The four sortable ones read from `leadershipListSortLabels` rather than
repeating the string, which is the treatment B17 already gave municípios and
dobradinhas — the picker's label and the sort summary cannot drift now.
The picker shipped as a capability of CampaignTable rather than a municípios
feature, so every list that already uses the shared table got it: pass
`columnVisibility` and the table mounts the menu and filters the columns.
Persistence is a cookie, not localStorage, and the reason is a type rather
than a preference: CampaignTableColumn.cell is a function, so the column array
never crosses the server/client boundary and the client cannot rebuild the
table without it. The server filters, so a hidden column does not reach the RSC
payload either. `campaign_columns` carries a listId -> hidden ids map in a
compact form with no JSON and no percent-encoding, because document.cookie and
Next's cookies() disagree about decoding.
Closing the menu is the commit. The first cut debounced at 400 ms, which
batches a burst rather than a reading pace, so unchecking four columns while
reading their labels was four full-route refreshes — on /campanha/liderancas
each one re-pays the ~229 ms / 371-statement render already ledgered P1. The
timer survives at 3 s as a safety net for a menu left open; the price is that
the table behind the menu no longer updates per checkbox.
`defaultVisible` is gone: it was true on 10 columns and false on none, and
storing only the hidden set is what makes a newly shipped column visible to
everyone who never touched the picker. In exchange `label` is required, which
made `head` optional — 24 columns had a head that only restated the new label.
Each surface declares one mandatory column, re-filtered on the server so a
stale cookie cannot leave a row without its only linked cell.
The bug review caught is this codebase's recurring one: the RSC payload of
toggle A arrives inside the window of toggle B, and adopting the server there
undid B and then committed the set without it. Reconciliation now runs only
when no write is pending, pinned red against the previous code.
Municipality column ids are typed against MunicipalityListColumnId now, so the
id, the B22 description and the picker label are one key; the union member
`nivel` was renamed `level` to match the id and the ?level= filter. The
picker's own content comparison is sameIdSet, widened to string ids.
