diff --git a/.github/playwright/README.md b/.github/playwright/README.md
index 0a34d24d5eef..2aa2b0107a77 100644
--- a/.github/playwright/README.md
+++ b/.github/playwright/README.md
@@ -17,7 +17,7 @@ SSO stays in its dedicated workflow, while knowledge graph and ontology share on
 ceil(total weighted worker time / (3 workers * 21 minutes * 0.85))
 ```
 
-The common matrix is bounded to 5–24 runners and uses a 21-minute allocation budget. Dedicated lanes use a 20-minute allocation budget. The common lane previously sat a minute below the dedicated lanes, but the chromium suite outgrew what 24 runners could hold at 19 minutes, so full-mode planning aborted outright; 21 minutes restores headroom while staying inside the 25-minute `timeout` wrapper around `npx playwright test` and the 35-minute `playwright-ci` job clock. Note that the allocation budget bounds a whole *shard*, whereas the strict 20-minute ceiling below bounds a single *atomic unit*, so the two are independent. Planner weights use the hook-inclusive observed duration, including retries. Only an exact stable test ID explicitly reported as skipped may retain a zero weight; every other zero-duration observation and every unseen test uses the conservative fallback, and zero weights never transfer through the file/title identity fallback. The versioned bootstrap baseline uses stable expected and skipped observations from the coverage-complete but failed full run `29984209316`, while unexpected and flaky tests retain their prior duration weights from run `29980474263`. It is bootstrap data, not a fabricated successful history; normal planning still uses p75 from the latest three successful full runs when those artifacts exist. This keeps expensive internally parallel suites together instead of multiplying their shared setup across runners. Serial/global behavior stays in one-worker lanes. Large suites listed in `AUDITED_PARALLEL_SUITES` are split at test granularity only after confirming that they are not serial and do not depend on earlier tests. The planner fails when any remaining atomic unit or bounded lane exceeds the 20-minute ceiling.
+The common matrix is bounded to 5–24 runners and uses a 21-minute allocation budget. Dedicated lanes use a 20-minute allocation budget. The common lane previously sat a minute below the dedicated lanes, but the chromium suite outgrew what 24 runners could hold at 19 minutes, so full-mode planning aborted outright; 21 minutes restores headroom while staying inside the `timeout` wrapper around `npx playwright test` (in `playwright-e2e-reusable.yml`) and the `playwright-ci` job clock. Note that the allocation budget bounds a whole *shard*, whereas the strict 20-minute ceiling below bounds a single *atomic unit*, so the two are independent. Planner weights use the hook-inclusive observed duration, including retries. Only an exact stable test ID explicitly reported as skipped may retain a zero weight; every other zero-duration observation and every unseen test uses the conservative fallback, and zero weights never transfer through the file/title identity fallback. The versioned bootstrap baseline uses stable expected and skipped observations from the coverage-complete but failed full run `29984209316`, while unexpected and flaky tests retain their prior duration weights from run `29980474263`. It is bootstrap data, not a fabricated successful history; normal planning still uses p75 from the latest three successful full runs when those artifacts exist. This keeps expensive internally parallel suites together instead of multiplying their shared setup across runners. Serial/global behavior stays in one-worker lanes. Large suites listed in `AUDITED_PARALLEL_SUITES` are split at test granularity only after confirming that they are not serial and do not depend on earlier tests. The planner fails when any remaining atomic unit or bounded lane exceeds the 20-minute ceiling.
 
 The `Basic` and `chromium` projects share that common 24-runner cap and are balanced together; they are not separate pools of standard hosted runners. Isolated ingestion, search, reindex, permission, and global-state lanes are additional because they cannot safely share mutable server state with the common matrix.