@@ -53,7 +53,7 @@ on:
   # A `pull_request` or `merge_group` run carries the FAST SET: the `test`
   # matrix below runs the test files the change affects, selected by
   # `scripts/select-affected-tests.mjs`. `push` to `main` and this hourly
-  # `schedule` carry the FULL BATTERY: every test file in twelve shards, plus
+  # `schedule` carry the FULL BATTERY: every test file in sixteen shards, plus
   # the eight-shard coverage lane. Nothing is deleted or weakened; the whole
   # suite moved to `main`, where it runs on every merge and at least hourly.
   # GitHub runs `schedule` on the default branch only. The minute is off the
@@ -921,8 +921,9 @@ jobs:
           path: .turbo/cache
           key: turbo-${{ runner.os }}-type-check-${{ github.sha }}
 
-  # PRs run the suite split across 8 runners. The suite is dominated by fixed
-  # per-file cost rather than by the assertions: a PR run reported 495s wall
+  # PRs run the suite split across the runners of the `shard:` matrix below.
+  # The suite is dominated by fixed per-file cost rather than by the
+  # assertions: a PR run reported 495s wall
   # clock of which only 178s was `tests` — the rest was `setup` (831s
   # cumulative), `import` (199s) and `environment` (156s), each paid once per
   # test file because `isolate: true` re-executes the module graph per file
@@ -960,7 +961,7 @@ jobs:
   # over half again as many runners. Duration-balanced sharding is that
   # card's second pull request, not this edit. The width is one number on
   # both speeds (objectui#12214): the affected run on a pull request or queue
-  # build and the full battery on `push` and `schedule` both use these twelve
+  # build and the full battery on `push` and `schedule` both use these same
   # legs, so `test-aggregate`'s `--shards` has one value to agree with.
   #
   # ⭐ Duration-balanced (objectui#12208, the same ruling's second pull
@@ -969,8 +970,25 @@ jobs:
   # lucky-heavy leg is the one the cap cancels. The note on the `Run tests`
   # step below says what replaced it; `test-aggregate` prints the longest leg's
   # headroom on every run.
+  #
+  # ⭐ 12 -> 16 (objectui#12384; the same ruling's width rule, "the width
+  # derived from the total measured duration at a target of about 10 minutes
+  # per shard"). Balanced, the twelve legs were no longer heavy -- measured on
+  # that card over the 50 hours after the balanced plan landed, 11.3 min
+  # median and 12.6 min p99 per finished leg -- but the runners under them
+  # were not all equally fast: the cap cancelled legs in 12 of the 279 runs
+  # that ran tests, with no hang, every file on those VMs running 1.6x to 2.0x
+  # slower than the same files on a normal one, episodically (13 of the 16
+  # cancelled legs fell on one afternoon, 2026-10-14) and with no runner label
+  # that tells such a VM apart. Twelve legs left a p99 leg 1.59x (20 / 12.6)
+  # of slack, below that factor; sixteen legs carry 0.75 of the load each,
+  # which moves the break-even slow factor to about 2.1x. The same three
+  # refusals hold: ⛔ the ceiling is NOT raised, ⛔ no test is skipped, and ⛔
+  # the relevance gate is not widened. A runner pin was not taken: there was
+  # no measured runner class to pin. Whether the width still fits is
+  # `test-aggregate`'s headroom reading on every run, not this comment.
   test:
-    name: Test (shard ${{ matrix.shard }}/12)
+    name: Test (shard ${{ matrix.shard }}/16)
     # Coverage instrumentation (v8 adds 40-100% overhead) is skipped on PRs;
     # the `test-coverage` job below runs the coverage lane on push.
     #
@@ -979,8 +997,9 @@ jobs:
     # that runs no tests is the hole this repository just paid for. `push` is
     # still excluded because the coverage lane below is the push lane: it is
     # sharded of its own since objectui#5403 (4 ways, 8 since objectui#12256) --
-    # ⛔ INDEPENDENTLY of the matrix above, which objectui#9499 widened to 8 and
-    # objectui#12208 to 12 without touching it: the two lanes declare their own
+    # ⛔ INDEPENDENTLY of the matrix above, which objectui#9499 widened to 8,
+    # objectui#12208 to 12 and objectui#12384 to 16 without touching it: the
+    # two lanes declare their own
     # `shard:` matrix and their own `--shard=N/M`, and the coverage blobs are
     # named `blob-N-M.json` by the coverage job's own flag -- and its
     # `coverage-report` job merges every shard's blob report into ONE complete
@@ -1005,7 +1024,7 @@ jobs:
       # Report every shard's failures, not just the first one to break.
       fail-fast: false
       matrix:
-        shard: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]
+        shard: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]
 
     steps:
       - name: Checkout code
@@ -1188,7 +1207,7 @@ jobs:
             merge_group) base="$QUEUE_BASE" ;;
             *)
               echo 'mode=full' >> "$GITHUB_OUTPUT"
-              echo "A $EVENT run is the full battery: every test file, twelve shards."
+              echo "A $EVENT run is the full battery: every test file, sixteen shards."
               exit 0
               ;;
           esac
@@ -1235,18 +1254,18 @@ jobs:
       # `scripts/__tests__/test-file-timings.test.ts` pins all three. The
       # writer never throws into the run: no report is one stderr line, never
       # a red leg.
-      - name: Run tests (shard ${{ matrix.shard }}/12)
+      - name: Run tests (shard ${{ matrix.shard }}/16)
         if: steps.relevant.outputs.should_run == 'true'
         env:
           MODE: ${{ steps.select.outputs.mode }}
           OBJECTUI_TEST_TIMINGS_OUT: ${{ runner.temp }}/test-file-timings/shard-${{ matrix.shard }}.json
         run: |
           if [ "$MODE" = 'mode=affected' ]; then
             mapfile -t filters < "$RUNNER_TEMP/affected-tests.txt"
-            echo "Affected speed: ${#filters[@]} vitest filter(s), shard ${{ matrix.shard }} of 12."
-            pnpm exec vitest run --shard=${{ matrix.shard }}/12 "${filters[@]}"
+            echo "Affected speed: ${#filters[@]} vitest filter(s), shard ${{ matrix.shard }} of 16."
+            pnpm exec vitest run --shard=${{ matrix.shard }}/16 "${filters[@]}"
           else
-            pnpm test --shard=${{ matrix.shard }}/12 --reporter=default --reporter=github-actions --reporter=./scripts/test-file-timings.mjs
+            pnpm test --shard=${{ matrix.shard }}/16 --reporter=default --reporter=github-actions --reporter=./scripts/test-file-timings.mjs
           fi
 
       # The full battery's per-file timings, for the weekly refresh of
@@ -1269,7 +1288,7 @@ jobs:
   # ── The built-artifact pin lane, lifted out of the shard matrix ──────────
   # (objectui#7183 for the lane itself; objectui#9499 for this job.)
   #
-  # `pnpm test --shard=N/12` above resolves every workspace package to its `src`
+  # `pnpm test --shard=N/16` above resolves every workspace package to its `src`
   # -- that is what the root alias map is for -- so no test in it can observe
   # its own package's BUILT output. This is the only place a `dist`-importing
   # pin runs, and `turbo run test:dist` carries `dependsOn: ["build"]` for the
@@ -1398,7 +1417,7 @@ jobs:
   # ⚠️ `needs.test.result` ALONE WOULD BE A PHANTOM GATE, and that is why this
   # job runs a script rather than reading an expression. A matrix job's
   # `needs.<job>.result` is a single ROLLUP over every leg, and a rollup cannot
-  # distinguish "all twelve ran green" from "nine ran green and three were
+  # distinguish "all sixteen ran green" from "thirteen ran green and three were
   # SKIPPED". A shard dropped from the matrix, or given a job-level `if:` that
   # is false, would leave this context GREEN while a fraction of the suite never
   # ran -- the objectui#8857 / objectui#3523 shape, a check reporting success
@@ -1443,7 +1462,7 @@ jobs:
           TEST_MATRIX_ROLLUP: ${{ needs.test.result }}
         run: |
           echo "needs.test.result (the matrix rollup this gate deliberately does NOT trust) = $TEST_MATRIX_ROLLUP"
-          node scripts/check-test-shard-results.mjs --shards 12 --job-name 'Test'
+          node scripts/check-test-shard-results.mjs --shards 16 --job-name 'Test'
 
       # The dist-pin lane is NOT a matrix, so its `needs` result is that one
       # job's own conclusion and `skipped` is visible in it -- no API read is
@@ -1473,7 +1492,7 @@ jobs:
         continue-on-error: true
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: node scripts/shard-headroom.mjs --job 'Test (shard {n}/12)' --shards 12 --cap-minutes 20 --target-minutes 10
+        run: node scripts/shard-headroom.mjs --job 'Test (shard {n}/16)' --shards 16 --cap-minutes 20 --target-minutes 10
 
   # Push to main/develop: the coverage lane, sharded 4 ways with a blob-report
   # merge (objectui#5403). The comment that used to sit here declined to shard
 
@@ -398,7 +398,7 @@ jobs:
       #
       # This step is a SECOND signal, not the gate. The gate is
       # `scripts/__tests__/check-test-path-roots.test.ts`, which runs inside
-      # `Test (shard N/12)`. Those shards are not required contexts themselves
+      # `Test (shard N/16)`. Those shards are not required contexts themselves
       # since objectui#9499 — the `Test` aggregator reports for all of them, and
       # it is required and subscribes `merge_group` — so the verdict still
       # blocks a queue build, through one name instead of N. This step reaches the same verdict sooner
 
@@ -21,7 +21,7 @@
 # fetched "the newest timings artifact" at their own start could straddle a
 # refresh, and two legs on different weights run some files twice and others
 # never, with every leg green. The committed file also makes a partition move a
-# reviewed, reproducible event -- `pnpm test --shard=3/12` on a laptop runs the
+# reviewed, reproducible event -- `pnpm test --shard=3/16` on a laptop runs the
 # same files CI's leg 3 ran -- and the dataset is in the root config's import
 # closure, so the pull request that moves it runs the full suite on the new
 # split before it lands (`scripts/select-affected-tests.mjs`).
@@ -51,13 +51,22 @@
 #
 # `scripts/test-file-timings.mjs select`: the newest completed `CI` runs on
 # `main` from the full battery's events (`push` and the hourly `schedule`,
-# objectui#12214) whose twelve `Test (shard N/12)` legs all reported success
-# and whose twelve `test-file-timings-N` artifacts are all present; at most
+# objectui#12214) whose sixteen `Test (shard N/16)` legs all reported success
+# and whose sixteen `test-file-timings-N` artifacts are all present; at most
 # six, and a file keeps the median of the runs that ran it. When no run
 # qualifies -- before the first full battery has uploaded, or after a week of
 # red -- it prints NOT MEASURED, writes nothing and exits green, because
 # nothing failed; a NOT MEASURED that persists is the finding.
 #
+# The `--shards` it selects by is the `test` matrix's width, and it moves with
+# it: a run of another width names other legs (`Test (shard 1/12)` is not
+# `Test (shard 1/16)`), so it is never eligible, and a stale value here selects
+# nothing forever -- the silent NOT MEASURED above, not a red run. Right after
+# a widening, the old-width runs are skipped by name and the first full
+# battery at the new width that goes green is the first one taken; a refresh
+# that fires before it exists keeps the committed file, which the sequencer
+# reads per file and never by width.
+#
 # ## The pull-request leg
 #
 # A pull request that edits this workflow or the scripts it runs executes it
@@ -119,10 +128,10 @@ jobs:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         run: |
           status=0
-          node scripts/test-file-timings.mjs select --shards 12 --limit 6 > "$RUNNER_TEMP/runs.json" || status=$?
+          node scripts/test-file-timings.mjs select --shards 16 --limit 6 > "$RUNNER_TEMP/runs.json" || status=$?
           echo "status=$status" >> "$GITHUB_OUTPUT"
           if [ "$status" -eq 3 ]; then
-            echo "::warning title=Test shard timings::NOT MEASURED: no completed full-battery run on main has all twelve test shards green and all twelve timing artifacts. scripts/test-shard-timings.json is untouched."
+            echo "::warning title=Test shard timings::NOT MEASURED: no completed full-battery run on main has all sixteen test shards green and all sixteen timing artifacts. scripts/test-shard-timings.json is untouched."
             exit 0
           fi
           if [ "$status" -ne 0 ]; then
 
@@ -187,10 +187,10 @@ export const SchemaRenderer = ({ schema }: { schema: BaseSchema }) => {
 ```bash
 pnpm exec vitest run packages/<pkg>/src/<file>.test.ts   # 只跑一个文件
 pnpm exec vitest run packages/<pkg>/                     # 只跑一个包
-pnpm test                                                # 全量(CI 就是它,可加 --shard=1/12)
+pnpm test                                                # 全量(CI 就是它,可加 --shard=1/16)
 ```
 
-`--shard=K/N` 在本地也经根 config 的 `sequence.sequencer`(`scripts/test-shard-sequencer.mjs`)分腿:按 `scripts/test-shard-timings.json` 里实测的每文件耗时,不按路径哈希,开跑前打印一行 `test-shard-sequencer:` 读数。分配只由文件集、那份 timings 与 N 决定(`scripts/test-shard-partition.mjs` 文件头),所以 N 取 CI 某个 matrix 的宽度时,同一 checkout 上不带路径的 `pnpm test --shard=K/N` 跑的就是那个 matrix 全量运行时第 K 腿的文件;带了路径,文件集变了,分配也跟着变。N 以 `ci.yml` 为准 —— `test` 与 `test-coverage` 两个 matrix 各自声明宽度;示例里的 `12` 是写下时 `test` matrix 的读数,本文件不复核它(诫条 #9)。
+`--shard=K/N` 在本地也经根 config 的 `sequence.sequencer`(`scripts/test-shard-sequencer.mjs`)分腿:按 `scripts/test-shard-timings.json` 里实测的每文件耗时,不按路径哈希,开跑前打印一行 `test-shard-sequencer:` 读数。分配只由文件集、那份 timings 与 N 决定(`scripts/test-shard-partition.mjs` 文件头),所以 N 取 CI 某个 matrix 的宽度时,同一 checkout 上不带路径的 `pnpm test --shard=K/N` 跑的就是那个 matrix 全量运行时第 K 腿的文件;带了路径,文件集变了,分配也跟着变。N 以 `ci.yml` 为准 —— `test` 与 `test-coverage` 两个 matrix 各自声明宽度;示例里的 `16` 是写下时 `test` matrix 的读数,本文件不复核它(诫条 #9)。
 
 AGENTS.md 的「只跑受影响的包」指的是**用上面的路径过滤缩小范围**,不是 `cd` 进包里。
 `pnpm --filter <pkg> test` 与 `turbo run test` **不再静默假绿**(objectui#3240):每个包的
 
@@ -1898,7 +1898,7 @@ describe('main', () => {
    * purpose. `docs-route-eager-closure.yml` and `performance-budget.yml` are
    * not among this repo's required merge-queue contexts, so a regression in
    * the fold would not block a merge through the gate's own job. This file
-   * runs inside `Test (shard N/12)`, whose verdict is required through the
+   * runs inside `Test (shard N/16)`, whose verdict is required through the
    * `Test` aggregator since objectui#9499 (objectui#9098 landed
    * the same reasoning one card earlier).
    */
 
@@ -51,7 +51,7 @@ import {
  *
  * Everything here runs on an UNBUILT tree: the fence scanner, the marker
  * convention, the JSON dialects, and the wiring. That boundary is not tidiness —
- * this suite runs inside `ci.yml`'s `Test (shard N/12)` jobs, which do not build
+ * this suite runs inside `ci.yml`'s `Test (shard N/16)` jobs, which do not build
  * the workspace, and a test that needed `dist/*.d.ts` would either be flaky or
  * would quietly assert nothing there.
  *
 
@@ -4,7 +4,7 @@
  *
  * ## Why the enforcement lives HERE and not only in a workflow step
  *
- * Every test file under `scripts` runs inside `Test (shard N/12)`, reported by
+ * Every test file under `scripts` runs inside `Test (shard N/16)`, reported by
  * the `Test` aggregator (objectui#9499), which is a
  * REQUIRED context and subscribes `merge_group` — so the scan below is what
  * actually blocks a queue build. The `Lint` step added alongside it is a second, faster
 
@@ -582,7 +582,7 @@ describe('the declared buckets partition what a pull request actually produces',
   it('found the workflows and the shard matrix (the parser still parses)', () => {
     expect(workflows.map((w) => w.file)).toContain('ci.yml');
     expect([...produced.keys()]).toEqual(
-      expect.arrayContaining(['Test (shard 1/12)', 'Test (shard 12/12)', 'Test', 'Lint', 'dependabot', 'label']),
+      expect.arrayContaining(['Test (shard 1/16)', 'Test (shard 16/16)', 'Test', 'Lint', 'dependabot', 'label']),
     );
     // No unexpanded template survived the matrix expansion.
     for (const name of produced.keys()) expect(name).not.toContain('${{');
 
@@ -344,8 +344,8 @@ describe('ci.yml wires the selector fail-open (objectui#12214)', () => {
   });
 
   it('keeps the matrix width and the full-suite command for the full battery', () => {
-    const run = code.slice(code.indexOf('- name: Run tests (shard ${{ matrix.shard }}/12)'));
-    expect(run).toContain('pnpm test --shard=${{ matrix.shard }}/12');
-    expect(run).toContain('pnpm exec vitest run --shard=${{ matrix.shard }}/12 "${filters[@]}"');
+    const run = code.slice(code.indexOf('- name: Run tests (shard ${{ matrix.shard }}/16)'));
+    expect(run).toContain('pnpm test --shard=${{ matrix.shard }}/16');
+    expect(run).toContain('pnpm exec vitest run --shard=${{ matrix.shard }}/16 "${filters[@]}"');
   });
 });
Original file line number	Diff line number	Diff line change
`@@ -398,7 +398,7 @@ jobs:`
`398`	`398`	`#`
`399`	`399`	`# This step is a SECOND signal, not the gate. The gate is`
`400`	`400`	# `scripts/__tests__/check-test-path-roots.test.ts`, which runs inside
`401`		- # `Test (shard N/12)`. Those shards are not required contexts themselves
	`401`	+ # `Test (shard N/16)`. Those shards are not required contexts themselves
`402`	`402`	# since objectui#9499 — the `Test` aggregator reports for all of them, and
`403`	`403`	# it is required and subscribes `merge_group` — so the verdict still
`404`	`404`	`# blocks a queue build, through one name instead of N. This step reaches the same verdict sooner`
Original file line number	Diff line number	Diff line change
`@@ -51,7 +51,7 @@ import {`
`51`	`51`	`*`
`52`	`52`	`* Everything here runs on an UNBUILT tree: the fence scanner, the marker`
`53`	`53`	`* convention, the JSON dialects, and the wiring. That boundary is not tidiness —`
`54`		- * this suite runs inside `ci.yml`'s `Test (shard N/12)` jobs, which do not build
	`54`	+ * this suite runs inside `ci.yml`'s `Test (shard N/16)` jobs, which do not build
`55`	`55`	* the workspace, and a test that needed `dist/*.d.ts` would either be flaky or
`56`	`56`	`* would quietly assert nothing there.`
`57`	`57`	`*`
Original file line number	Diff line number	Diff line change
`@@ -4,7 +4,7 @@`
`4`	`4`	`*`
`5`	`5`	`* ## Why the enforcement lives HERE and not only in a workflow step`
`6`	`6`	`*`
`7`		- * Every test file under `scripts` runs inside `Test (shard N/12)`, reported by
	`7`	+ * Every test file under `scripts` runs inside `Test (shard N/16)`, reported by
`8`	`8`	* the `Test` aggregator (objectui#9499), which is a
`9`	`9`	* REQUIRED context and subscribes `merge_group` — so the scan below is what
`10`	`10`	* actually blocks a queue build. The `Lint` step added alongside it is a second, faster