Catalog: retire the last holding dropdowns, and take the published packages

[JustinSGray]

Addresses APP-224.

Brings paul/tool_catalog up to date with main and closes out the tool
assembly tree's last loose end. The branch also stops depending on a local
checkout, so it installs in CI for the first time.

The last two dropdowns

The assembly tree took over filling a holder or collet slot on 2026-09-08, and
main removed the dropdowns from the tool table and the panel on 2026-09-10.
What neither took out was the machinery on the route that had fed them, which
outlived them because nothing on screen was drawn from it any more.

picked — a map of holder and collet guids keyed by tool — had no writer left,
so it could only ever be empty, and an empty map still read like an unsaved
answer at three call sites. Gone, and with it:

• pick and saveAssembly, whose only caller was the panel, and the per-tool
  optionsFor cache behind them.
• panelActions's assemblyChanged, which compared picked against the
  ordered line to offer Update. Reading the empty pick as a change would have
  offered Update on every tool ordered with a holder. It is now false.

The Tool / Tool + holder switch goes too, for the reason the dropdowns did:
it was a second answer to a question the tree settles, and it let the one
picture on the page draw a bare cutter beside a stack the tree had fully
assembled. What a reader wanted from the Tool half was the working end drawn
bigger, so the press is now a zoom — @toolpath/tool-drawing's zoom prop,
which cuts the sheet just above the holder nose rather than dropping the holder
out of the picture. LBH reads the stack's stickout under both framings.

Two debug modules

Both pure, both tested, both behind import.meta.env.DEV:

• clearance-debug.ts dumps everything the clearance wall was drawn from as a
  paste-able case. A wall that collapses against the cut looks identical whether
  the curve arrived in inches, the sheet reserved no room on the +r flank, or
  the staircase has no rise in it — three unrelated causes, one picture.
• holder-debug.ts reports which of five gates emptied a holder rack. "Why are
  there no shrink-fit holders for this feature" is a question the page cannot
  answer and this can.

Published packages, and no more link: overrides

The branch carried two pnpm.overrides link: entries onto a sibling
ui-packages checkout. Both are now obsolete:

• @toolpath/tool-drawing 0.3.2 → 1.0.0 (apps/catalog). This is what the
  branch was blocked on — Zoom, the zoom prop, DrawingContext.extent and
  Frame.reserve are all in 1.0.0; against 0.3.2 the catalog failed tsc with
  five errors naming them.
• @toolpath/tool-scraper 2.5.0 → 3.0.0 (packages/catalog-data).

Both are major bumps and neither needed a source change here. pnpm install
now resolves entirely off the registry, so --frozen-lockfile works on a clean
clone.

Checks

pnpm check is green — check-style, lint, build, check-types and
all 1344 tests, locally and in CI.

CI also gets through pnpm install --frozen-lockfile for the first time on this
branch. That step failed before, because the link: overrides pointed at a
checkout no runner has.

The materialRoom workaround, retired

tool-drawing 1.0.0 deliberately retired caller-supplied padding as a margin.
padding is now a reservation, granted out of room the drawing itself
cannot use — asking for more than the panel has costs the drawing nothing, and a
narrow sheet grants what it has. The guidance is to ask for as much as the
widest sheet could use.

That is one number rather than a per-caller one, so the workaround this
application carried for it comes off:

• materialRoom off CatalogDrawing, which now asks for MATERIAL_ROOM always.
• PANEL_MATERIAL_ROOM off tool-details.tsx. It existed only to dodge the 0.6
  clamp that crushed the assembly into the top third of the panel on 2026-09-03,
  and there is no clamp to dodge now.

The test inverted rather than being deleted. It pinned that asking for less
left the tool more of the sheet — the workaround's own symptom, and the one
assertion that failed on 1.0.0, because there is nothing left to trade. What it
pins now is the guarantee that made removing the prop safe: the only reservation
the component still varies is whether there is an overlay to reserve for at all,
and the tool has to be framed the same either way.

Verified as a sensor rather than a restatement — run against 0.3.2 it fails,
97.39 reserved against 90.55 bare. The reservation coming out of the drawing
is exactly the regression it exists to catch, and it is one this application
cannot see for itself.

Notes for review

• The merge from main had six conflicts, all where this branch's dropdown
  removal met the same removal done independently on main. main's side won
  wherever it had already done the work, including its stronger column test.
  One of its tests needed adapting: it clicked the Tool switch to assert the
  stickout fell back to the tool's own LBH, and there is no bare-tool view to
  fall back to — it now presses the zoom and asserts the number does not move.
• OrderDialog still carries a holder and collet pair of its own, left standing
  deliberately: it has been unreachable since before any of this (setAdding is
  only ever called with null), so removing it is a separate decision.
• @toolpath/api (0.4.1 declared, 0.6.0 published) and @toolpath/tool-support
  (0.3.0 declared, 0.3.1 published) are still behind. Out of scope here.