/system lists each first-opened page's warm median beside its budget, one plain line per page #11567

/system now has a section, "First-opened pages against their budget", with one line per page in PAGE_BUDGETS order.

An over-budget page reads "map answers in 82 ms, budget 50 ms". A page inside its budget reads "… within its N ms budget". A page nobody opened reads "not opened in the last 15 minutes (budget N ms)". There is no banner. The lines sit behind one <details> summary ("Page budgets: all met" or "N over"), so the first view stays in budget.

The gate, design-lint and tsc --noEmit came back clean. The new and touched tests and test/mcp-size.test.ts pass (12 pass, 0 fail). The branch was already pushed and git push reported nothing new, so the forge's check on this head is the one to watch. I did not boot the app or open /system.

• Rows: src/perf/page_budget_read.ts has pageBudgetRows(samples, now), a pure function. It takes samples keyed by router pattern from the existing perf record, drops anything older than the 15-minute window, and returns { page, pattern, budgetMs, statements, medianMs | null, n, over }. over is true only when the median is past the budget.
• No new SQL: the read is in memory only, so /system's statement count does not rise. src/web/pages/system.ts mounts the section once.
• Files outside the listed ones: I removed src/perf/page_budget.ts from scripts/unreached-allow.json, since the new code makes it reachable. I also added the new section's heading to test/system-parts.test.ts, which counts one h2 per part.
• Tests and docs: test/page-budget-read.test.ts covers the row order, null for unsampled pages, the over-budget boundary and the render. docs/features/page-budget.md is one short paragraph.

I left out any red styling, alerts or history of past medians. I did not touch the MCP.

Closes #11544

/system now has a section, "First-opened pages against their budget", with one line per page in PAGE_BUDGETS order.

An over-budget page reads "map answers in 82 ms, budget 50 ms". A page inside its budget reads "… within its N ms budget". A page nobody opened reads "not opened in the last 15 minutes (budget N ms)". There is no banner. The lines sit behind one <details> summary ("Page budgets: all met" or "N over"), so the first view stays in budget.

The gate, design-lint and tsc --noEmit came back clean. The new and touched tests and test/mcp-size.test.ts pass (12 pass, 0 fail). The branch was already pushed and git push reported nothing new, so the forge's check on this head is the one to watch. I did not boot the app or open /system.

• Rows: src/perf/page_budget_read.ts has pageBudgetRows(samples, now), a pure function. It takes samples keyed by router pattern from the existing perf record, drops anything older than the 15-minute window, and returns { page, pattern, budgetMs, statements, medianMs | null, n, over }. over is true only when the median is past the budget.
• No new SQL: the read is in memory only, so /system's statement count does not rise. src/web/pages/system.ts mounts the section once.
• Files outside the listed ones: I removed src/perf/page_budget.ts from scripts/unreached-allow.json, since the new code makes it reachable. I also added the new section's heading to test/system-parts.test.ts, which counts one h2 per part.
• Tests and docs: test/page-budget-read.test.ts covers the row order, null for unsampled pages, the over-budget boundary and the render. docs/features/page-budget.md is one short paragraph.

I left out any red styling, alerts or history of past medians. I did not touch the MCP.

I found nothing in this change that could cause that failure, so I merged current main and pushed again. I have not seen a passing check yet.

• What the log shows: every test in the tail passed (112 pass, 0 fail). The failure came from the machine, not a test. A Chrome process, which the browser-based fit tests start, tried to reach Google hosts the repository's egress refuses. The shards then used 16.7 cores for 30s against a limit of 16, and the check was stopped.
• Why I think it isn't this change: the pull adds one pure function, one render module and one mount on /system, and none of it starts a process or a browser. The overload was a CPU limit on the shared check machine, so it may come back whatever this pull does.
• What I pushed: I merged the 17 commits main had gained, which may include relevant fixes, and pushed 781f1c296. Locally the gate and tsc --noEmit came back clean, and test/page-budget-read.test.ts and test/system-parts.test.ts pass (8 pass, 0 fail). I did not run the browser-based tests here.

If the next check stops on the same limit, a person should look at the browser tests' CPU use or the shard count, because that is outside this pull's files.

The page-budget section is done and pushed, but the check has now failed twice on the machine's CPU limit and I have not changed anything to fix it. I found no cause in this pull.

Both failures are the same: the check was stopped for using about 17 cores for 30s against a limit of 16, with sh as the busiest process. Every test that ran passed (112 pass, 0 fail). Nothing in this pull starts a process or a browser. The only browser-driven test in the stopped shard, test/people-fit.test.ts (with test/helpers/chrome.ts), dates from earlier pulls (#9399 and #9079). The log shows Chrome trying to reach Google hosts that the repository's egress refuses. A busy shared machine is my best guess. I have not confirmed it, and I don't know whether other pulls hit the same limit. A person should look at the browser-test shards' CPU use. Adding a commit to this pull would not change that.

What changed

/system has a new section, "First-opened pages against their budget", with one plain line per page in PAGE_BUDGETS order. An over-budget page reads "map answers in 82 ms, budget 50 ms", with no banner. A page nobody opened reads "not opened in the last 15 minutes". The lines sit behind one summary ("Page budgets: all met" or "N over"). The read is in memory from the perf record, with no new SQL.

What to look at

• src/perf/page_budget_read.ts holds the pure pageBudgetRows(samples, now).
• src/web/pages/system-budget.ts renders the section, and src/web/pages/system.ts mounts it once.
• test/page-budget-read.test.ts covers row order, null for unsampled pages, over only past the budget, and the render.
• scripts/unreached-allow.json no longer lists src/perf/page_budget.ts, now that it is reachable.
• test/system-parts.test.ts gains the new heading.
• docs/features/page-budget.md is one short paragraph.

Left out

Red styling, alerts, history of past medians, and the MCP.

Checked locally

The gate, bun x tsc --noEmit, design-lint and the touched tests are clean, and test/mcp-size.test.ts passes.

To try it

Open /system on the preview and find the "Page budgets" summary.

Parked by the crew: the check stayed red through 2 automatic tries: check failed: stopped: used 17.0 cores for 31s, over its limit of 16; the busiest process is sh (pid 808602, 462s of CPU)