/forges: an owner's forges side by side, with size, use against size, apps and last heartbeat #6579

/forges now opens with a "Forges side by side" table above the existing client list. I ran bun x tsc --noEmit, design-lint and the tests for what I touched, and they pass. I did not run the whole check, and I did not boot the app or look at the page. The new tests don't render the table.

Each forge gets one row with the same columns, so nothing shifts: name, address, size, memory, CPU and disk, included apps, and the last heartbeat received.

• Hosted child: memory, CPU and disk read "shared".
• Child in its own cell: it shows use against its size, e.g. "310 MiB of 512 MiB". If the cell isn't running it says "not running" with the size's figures.
• No heartbeat yet: the cell reads "none yet".

Where it's wired in:

• Cell reading: it comes from the existing cell watch in src/apps/contain.ts. I added a CPU figure to it, averaged over the watch's window, and it stays empty until the window has half filled.
• New module: src/spawn/usage.ts turns a child's reading into plain data and words.
• Row fields: size, use and lastHeartbeatAt are now on each client row. The forges MCP tool and the terminal screen read those rows, so an agent gets the same fields.
• Docs and tests: docs/features/forges-side-by-side.md describes it, and test/forges-side-by-side.test.ts covers the reading and the not-found case.

What I left out or chose:

• Access: only the forge's admins see /forges, as before. The ledger doesn't record a separate per-forge owner, so "owner" means this forge's admin. A non-admin still gets not_found. If you want owners who aren't admins, the ledger would need to record who sent each forge.
• Disk: it's a bounded walk of the child's data directory, kept for 30 seconds. It is not a live disk meter.
• Layout: I kept the existing row list below the table instead of replacing it, so nothing that used its rows breaks.

Closes #2933

/forges now opens with a "Forges side by side" table above the existing client list. I ran bun x tsc --noEmit, design-lint and the tests for what I touched, and they pass. I did not run the whole check, and I did not boot the app or look at the page. The new tests don't render the table.

Each forge gets one row with the same columns, so nothing shifts: name, address, size, memory, CPU and disk, included apps, and the last heartbeat received.

• Hosted child: memory, CPU and disk read "shared".
• Child in its own cell: it shows use against its size, e.g. "310 MiB of 512 MiB". If the cell isn't running it says "not running" with the size's figures.
• No heartbeat yet: the cell reads "none yet".

Where it's wired in:

• Cell reading: it comes from the existing cell watch in src/apps/contain.ts. I added a CPU figure to it, averaged over the watch's window, and it stays empty until the window has half filled.
• New module: src/spawn/usage.ts turns a child's reading into plain data and words.
• Row fields: size, use and lastHeartbeatAt are now on each client row. The forges MCP tool and the terminal screen read those rows, so an agent gets the same fields.
• Docs and tests: docs/features/forges-side-by-side.md describes it, and test/forges-side-by-side.test.ts covers the reading and the not-found case.

What I left out or chose:

• Access: only the forge's admins see /forges, as before. The ledger doesn't record a separate per-forge owner, so "owner" means this forge's admin. A non-admin still gets not_found. If you want owners who aren't admins, the ledger would need to record who sent each forge.
• Disk: it's a bounded walk of the child's data directory, kept for 30 seconds. It is not a live disk meter.
• Layout: I kept the existing row list below the table instead of replacing it, so nothing that used its rows breaks.

The four client-list tests failed because of my change. My new table marked its rows data-forge="<name>", the same attribute the existing list rows use. The tests look a row up by that attribute, so they found my table row first and read the wrong markup.

I renamed the table's attribute to data-compare. test/client-list.test.ts, test/forges-side-by-side.test.ts and test/forges-timing.test.ts now pass (17 tests), and bun x tsc --noEmit is clean. I pushed that fix; I haven't seen the forge's result for it yet.

Nothing else changed: /forges still opens with the side-by-side table above the existing client list.

I've addressed everything from the review except the 360px look. Tests pass locally, and bun x tsc --noEmit and design-lint are clean, but I haven't seen the forge's check on this push.

• Terminal: each line in the "Forges you sent" card now ends with size, use and heartbeat. A hosted child reads small · shared, and a cell in use reads memory 310 MiB of 512 MiB · cpu 0.4 of 1 core · disk 100 MiB of 2 GiB. A forge with no heartbeat reads no heartbeat yet. I updated three expected strings in test/tui-forges.test.ts because the lines are longer now; the old strings were right for the old shape. A new test in that file covers shared, cell and idle.
• Table tests: test/client-list.test.ts now checks that an admin gets 200 with all eight columns, a shared row and none yet. test/forges-side-by-side.test.ts checks the not-running and X of Y cells. It also checks coresOf, which I made a pure exported function: it returns null before half the window and a figure after.
• Disk: I pruned the cache, so a reading nobody has asked for in 5 minutes is dropped. I did not take disk off the page request. The walk is still synchronous, bounded, and runs at most once per cell child every 30 seconds. Reading it from the watch would mean changing the watch to walk each cell's data directory itself, which I judged larger than this pull.
• Two lists: the table is for comparing and the list below it stays for actions (claim state, fresh link, spend, next step). This is written in docs/features/forges-side-by-side.md. I kept both because the actions don't fit table columns.
• Idle cells: a cell that isn't running reads not running under Memory, CPU and Disk. The Size column now carries the limits for every row, e.g. small · 512 MiB, 1 core, 2 GiB.
• 360px not checked: I can't run a browser here, and an anonymous request to the preview gets the 404 you saw. The table sits in the existing table-scroll wrapper, and the forge name is a row header. Someone with an admin session should confirm the scroll and that the first column stays readable.

Both fixes and the two smaller ones are pushed. I ran the tests for what I touched, bun x tsc --noEmit and design-lint, and all pass. I haven't seen the forge's check on this push, and I still haven't looked at the table in a browser.

• Scroll region: the table's wrapper is now role="region" tabindex="0" aria-label="Forges side by side", and the table has data-sticky-lead, so the Forge name stays put while the figures scroll. A test checks these attributes.
• Disk off the request: /forges no longer walks a directory. It reads the last cached figure and, if that is missing or older than 30 seconds, schedules a refresh on a timer after it has answered. So a cold first view shows measuring · 2 GiB in the Disk column, and the next view has the number.
• Walk still blocks: the refresh is still a synchronous walk of a bounded directory, so it blocks the event loop briefly when it runs, just not inside the request. I did not move it to the watch tick or the heartbeat handler. That would put the walk on a schedule even when nobody opens the page, and I judged that worse than walking only on demand.
• Unknown disk: diskMiB is null, not 0, when the figure isn't known yet or the directory is unknown. The docs say so.
• README: one line added about the table.