chore(ci): give browser proof headroom with fresh weights, a 20-minute bound and three shards

.github/workflows/ci.yml

@@ -302,11 +302,11 @@ jobs:
     needs: changeset
     if: needs.changeset.outputs.browser == 'true' && needs.changeset.outputs.run_browser == 'true'
     runs-on: ubuntu-24.04
-    timeout-minutes: 15
+    timeout-minutes: 20
     strategy:
       fail-fast: false
       matrix:
-        shard: [1, 2]
+        shard: [1, 2, 3]
     steps:
       - uses: actions/checkout@v7
         with:
@@ -323,7 +323,7 @@ jobs:
           HEAD_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
         run: bun scripts/ci-plan.ts
       - name: Complete browser proof
-        run: bun scripts/ci-proof.ts browser --shard=${{ matrix.shard }}/2
+        run: bun scripts/ci-proof.ts browser --shard=${{ matrix.shard }}/3
       - name: Proof receipts and failure diagnostics
         if: always()
         uses: actions/upload-artifact@v7

CONTRIBUTING.md

@@ -272,7 +272,7 @@ feature test should use these helpers instead of building its own.
 
 The complete dictation browser fixture has its own `dictation` project so its
 Chromium touch emulation and consumer styles stay outside the shared `visual`
-page. The default browser wrapper includes it in both shards; use
+page. The default browser wrapper includes it in every shard; use
 `node scripts/visual.mjs --project=dictation` for a focused container run.
 
 Tests and typecheck run from live TS source — no build needed. The

docs/decisions/ci-utility.md

@@ -111,7 +111,8 @@ one PR share a cancellable group. Different PRs never cancel one another.
 
 After cheap gates, packaging, strict verification, unit shards, runtime probes,
 browser shards, layout shards, a separate endurance job and captures start independently.
-There are four measured-cost unit partitions and two browser/layout partitions,
+There are four measured-cost unit partitions, three browser and two layout partitions
+(`PROOF_SHARDS`, `scripts/proof-shards.ts`, which the workflow guard matches against the matrices),
 with full coverage assertions and original per-browser file concurrency bounds.
 The first two-unit-partition hosted run reached its 20-minute budgets while
 still making progress. Completed-file observations in
@@ -141,7 +142,10 @@ inputs, never runtime dependencies. Layout refresh admits ordinary
 fresh job and never becomes an ordinary browser/layout weight. Requalify timing
 inputs after scenario ownership changes: the former combined recording-file
 measurement includes endurance and is not its current ordinary-layout cost.
-Two layout shards, per-project serial execution and all deadlines remain unchanged.
+Layout keeps two shards; per-project serial execution and all test deadlines remain unchanged.
+Browser proof runs in three shards and its jobs have a 20-minute bound. It was raised from 15 minutes after browser
+(2) reached 887 s on main and 917 s on a ready PR ([#2414](https://github.com/schlessera/brain-kit/issues/2414)). A weight refresh still
+qualifies by measured whole-job elapsed time, not by the bound.
 
 Hosted unit partitions use a queue capped at two serial Bun test processes
 (`scripts/unit-workers.ts`), or one when only one file is selected. Native
@@ -340,7 +344,7 @@ and actual merge-result parent/tree validation remain binding.
 
 ### Retaining complete browser domains
 
-Under #1333, ready-PR metadata can retain both shards of a complete browser or
+Under #1333, ready-PR metadata can retain every shard of a complete browser or
 layout domain from up to three recent successful automatic runs of that same
 contribution. Unit, types, packaging, native runtime and editorial checks remain
 fresh: their Git-sensitive or broader process inputs do not have an admitted
@@ -556,7 +560,7 @@ locked installed resources. No current output or untracked image enters this
 snapshot, and baseline changes cannot inherit old passing comparisons.
 
 The inclusion fingerprint binds relevant immutable Git paths, blob identities
-and modes, selected discovery ownership/order, the two measured partitions and
+and modes, selected discovery ownership/order, the domain's measured partitions (`PROOF_SHARDS`) and
 runner identity. It also includes every proof/tooling script and workflow,
 root/package manifests and lockfiles, compiler configuration and expanded build
 source inventories. These conservative shared harness inputs can invalidate

docs/process/proof-domains.md

@@ -77,8 +77,8 @@ Reconcile each final command as follows:
 | --- | --- | --- | --- |
 | Curated contract/integrity files in `FAST_TESTS` | Fast | Affected subset with strict types | Included once by complete affected unit discovery |
 | Every other product Bun test in `packages/` and `tests/` | Unit confirmation | Deferred | Complete affected discovery plus root tests |
-| Storybook dark/light, visual, pointer and isolated browser projects | Browser confirmation | Deferred | Every discovered project/file instance in the existing two shards |
-| `.layout.tsx` and `.offline.tsx` in `ui-react-layout` | Layout confirmation | Deferred | Every discovered project/file instance in the existing two shards |
+| Storybook dark/light, visual, pointer and isolated browser projects | Browser confirmation | Deferred | Every discovered project/file instance in the three browser shards |
+| `.layout.tsx` and `.offline.tsx` in `ui-react-layout` | Layout confirmation | Deferred | Every discovered project/file instance in the two layout shards |
 | `.endurance.tsx` in `ui-react-endurance` | Endurance confirmation | Deferred | Every discovered file in a separate concurrent job |
 | Native permission/enforcement/delegation/capability/host/cleanup probes | Runtime confirmation | Deferred | Existing complete command inventory and offline namespaces |
 | Packed consumer probes and editorial capture verification | Packaging/capture confirmation | Deferred | Existing affected commands, pins and artifacts |