AI Slop Removal - #49

`1"` tokenized to a bare `1`, which prefix-matches every 1.5", 10-32 and
16T in the library, so the sizes a user asked for were buried. The mark
now stays on the number and ends its token, leaving `1"` an exact size
while a bare `1` still prefixes them all.

Phosphor deprecated the unsuffixed icon names in favor of `*Icon`. The
local heart wrappers become FavoriteIcon/UnfavoriteIcon, which is what
they meant anyway and leaves HeartIcon to Phosphor.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

Searching a size found the element but showed its default configuration:
`1` prefix-matches a segment of every part number (0016 -> 16), which
tied the real match in the name, and the tie went to the part number's
first record. A term matched whole now outranks one matched as a prefix,
and a size's unit is folded into the mark, so `1`, `1"` and `1 in` all
name the 1" configuration rather than the .25" one listed first.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

CardTitle took three thumbnail props only to forward them, so the rows
that know which configuration to render now pass the CardThumbnail
itself. The part name and number block becomes its own component rather
than a nested ternary in the middle of the title.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

Every list narrows its insertables one of two ways and then renders the
same rows, but each caller resolved hits to insertables and tallied the
filtering itself (filterResult2 beside filterResult). Both ways now
return FilteredInsertables, and the favorites list splits into the
search, the browse, and the cards they share.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

Favoriting from the insert menu or a search result dropped the
configuration in front of the user, so the new favorite came back as the
element's default and had to be edited into the one just picked. The
button and menu item now carry it, canonical, the way the favorite menu
stores it.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

The app could only be signed in to: standalone there was no way to leave,
short of clearing cookies. Sign-out drops the session and the access
level resolved from it, then lands the caller back where they stood. Only
offered outside Onshape, whose panel session is Onshape's to end.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

The tab stays, so the library is announced, but selecting it lands on a
zero state instead of an empty library: its groups, favorites and search
all render inside the route, so gating there covers them at once, and
the loader stops fetching what nothing will show. The search row goes
with them, having nothing left to search.

Also says why the group-restore latch is there, which a reviewer had to
ask: without it, leaving a group redirects straight back into it.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

Where a caller left off was client state, so the entry redirect could
not see it and the app had to bounce itself into the group after
loading, guarded by a module-level latch against redirecting straight
back out. The group now sits in the user row beside the library it
belongs to, and entry computes the whole landing url in one query: the
join is also the check, so a deleted group, or one left behind by a
library switch, lands in the library instead.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

Six review notes, all pulling the same way — one local store the app
reads from, instead of state spread across a second localStorage key,
the url, and a hook returning a pair:

- ui-state gains useGetUiState/useSetUiState, so a component reaches for
  the half it wants instead of indexing a tuple.
- local-settings is gone: the caller's settings live in ui-state, which
  is now the source of truth for them. Saving still writes the row when
  they are signed in, which is what a browser running the app for the
  first time, and the Onshape launch, start from.
- The theme comes off ui-state rather than the url, so a standalone visit
  with no parameters still renders in it. The entry redirect seeds the
  account's theme once; /app takes it and drops the parameter, and the
  retained set is now the launch parameters by name.
- A sign-in leaves a flag in ui-state rather than a url marker, and
  returns to the entry point, which confirms it and resumes in one place.
- AppSelect and its option helpers had one caller, so they collapse into
  a SettingSelect beside it: values in, capitalized labels out, no memos.
- CHROME_BACKGROUND is jargon; it is the background of the bars framing
  the page, so FRAME_BACKGROUND.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

The mark still ends its number, which is what the tests that remain
cover; `1 in` still reaches the 1" configuration through the term
scoring rather than through canonicalization.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

Trimming and case-folding were sprinkled across the record parser, the
vendor lookup, the index and the frontend's display rule, and the
placeholder rule ran only at render — so `N/A` reached the index and
searching it returned noise, while "repeats the name" was written twice.

`clean` and `equalsIgnoreCase` are now a shared leaf, and one
`meaningfulPartNumber` says whether a number identifies anything. The
frontend's displayPartNumber delegates to it, so what is shown and what
is indexed cannot drift apart. The escaped-slash literals read as
strings while they are being touched.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

One tokenizer ran over every field, so a part number was read as prose:
217-2600 became 217 and 2600, TTB-0016 lost its zeros, and the fraction
in TTB-0016-5/32 was folded into a decimal. None of that identifies the
part any more.

Tokenizing is now field-aware. A part number is indexed as typed plus
its segments, so it is found whole or by either half, zeros included. A
name keeps the decimal canonicalization its sizes need, since the
standards write the same measurement as 1/2" and 0.5". A query has no
field, so it offers both readings — minus the pieces that would only
flood: a bare size is not split into digits, and the lone letters left
by splitting `n/a` are dropped. Commas no longer stick to the word
before them, which the dimensions in a name are full of.

The frontend scorer reads each field the same way, and the leading-zero
underline hack goes with the zeros it worked around.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

Some vendors genuinely sell parts without part numbers, so the warning
fired as noise on them. Checking the format of a number that is there is
the useful check, and that is a separate piece of work.

Stored rows keep the issue until a reload recomputes them, where it
renders as nothing at all rather than an error.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

The rename refactor replaced the resolution chain with an if-chain and
dropped its middle arm, so a record naming McMaster-Carr resolved to the
insertable's tagging instead — which its test says it should not. Put
the arm back, in the new shape.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

Dropping every one-letter term took the first keystroke of every search
with it, and answering as the caller types is most of what the search
is for. Only a letter left behind by splitting a longer word is dropped
now — the `n` and `a` of `n/a`, which was the case that wanted it.

The same measurement reaches the library written both ways: one vendor
writes .196 as .2 and the next writes .19. A name is stored and searched
as both spellings, so either finds the part. Most numbers spell the same
either way and cost nothing.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

The stray-letter rule was the wrong shape for what it was protecting
against. Nothing carries `n/a` — ingest drops it — but typing it split
into `n` and `a`, and a one-letter prefix answers with most of the
library. The placeholder is now recognized where it is typed, by the
same rule that drops it where it is stored, so it finds nothing and the
rest of the query is read as usual.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

The two branches were a 35-line ternary argument, which read as one
expression when it is really the same push either way.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

Seventeen icons were each wrapped in a Box to take a Mantine color, and
the colors themselves were written at every control: c="red" here,
color="yellow" there. AppIcon is that wrapper, and StatusColor names what
each color means, so an error looks like an error everywhere and the
theme's variables stay out of the markup.

The build-status icons collapse into one mapping from severity, since
their four branches differed only in the icon and the color. Elements
become parts in the group's zero state, which is what a library holds.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

A rate limit reached the client as its own kind carrying the seconds to
wait, which nothing on the client read: the loader is what waits, and all
the caller can do is try again. It is a handled message now.

401 and 403 take its place as kinds, because those do differ in what the
caller can do: a missing session is offered a sign-in in the toast
itself, where before both said "try signing in again" and left them to
find the button.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

Names and comments first: the private writer was `setUiState` beside the
public `useSetUiState`, the storage key was written twice, and both
accessors called themselves asynchronous while being neither.

The parse was the real problem. State stored by an older build, or naming
a library that no longer exists, threw inside getUiState — which routes
call before rendering, so the app would fail to start with nothing but
clearing site data to recover. It falls back to the defaults now: losing
a preference beats losing the app.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

- The frontend's part-number module re-exported one backend function
  under a second name; its two callers now import the rule itself.
- parseUrl parses an Onshape document url, so parseOnshapeUrl.
- The feedback form url sat in url.tsx; it belongs to the setting that
  opens it.
- The default access level is granted as well as viewed, so a dev
  override is not clamped back to user while access data is pending.
- RequireAccessLevel asked "is it admin, or is it editor" in two
  branches, which is isWithinAccessLevel — and answers a required USER
  level correctly, which the branches did not.
- useIsJobRunning is the read several places wanted, leaving the refresh
  on finish to its single owner.
- SeverityCounts is an interface rather than an inline shape.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

Canonicalizing a quantity rounded the parsed base value to seven decimals
of its own choosing and assembled the unit string by hand, beside a
module that already knows what a value is measured in and how close two
of them have to be to count as one. formatBaseValue is that spelling:
base units, to the decimals the tolerance distinguishes, so two values
the parser reads as equal cannot key two renders.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

The build status mapped a configuration row into a shape that differed
from it only by dropping the id, so the id joins the contract and the row
goes through as it is. Its private cache gets the reason written down:
the body is the same for every editor, but a shared cache would hand it
to whoever asked for the url next.

isSignedIn reads as the one question it answers, and clean stops saying
"blank is nothing" twice.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

`c`, `i` and `warm` become `configuration`, `insertableId` and
`renderThumbnail`, which is what the flag asks for: start the render this
url is missing. `v` stays, being the cache version every immutable url
carries, and the one stored urls already hold.

The stand-in served while a configuration renders was cached for a
minute, which is a minute of anyone asking that url getting the old
answer after the real render landed. It is a body the url does not
promise, so nothing stores it now: the next request is free to see past
it, and polling for a thumbnail we want costs an R2 read.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

Reaching for the middleware meant a context flag telling it to undo
itself, for the one route whose answers genuinely differ: stored bytes
the url pins, a stand-in it does not, and a miss that must stay
retryable. setCache says that per response, and the middleware goes back
to being for routes whose every answer is alike.

The query param is canonicalConfiguration, since that is what it is: a
url that took an ordinary configuration would key a different render.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

Every c.json in the error handler cast its status through never, because
ApiError carried a bare number and c.json takes the statuses a body is
allowed with. Carrying Hono's ContentfulStatusCode instead drops all four
casts and makes the constraint real: a 204 as an error status is now a
type error rather than a response the client cannot read.

The import is type-only, so the leaf the frontend imports stays free of
Hono in the bundle.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

A stored record keys on the canonical spelling of its selection, not on the
selection as written, so call the field canonicalConfiguration everywhere it is
stored or read — including in the D1 records column and the R2 search index.

The parse side still holds the selection as sent to Onshape, which only
becomes canonical in toResult, so it gets its own ProbedRecord.

Stored records only carry the new field after a library reload.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

Three things were called some variant of "configuration", and two of them
shared a name under different types: the selection a user made, that selection
normalized into a map, and the text form of that map addressing a render.

There is no normalized map any more. canonicalizeConfiguration returns the
text directly, so a canonicalConfiguration is always a string and a
configuration is always ParameterValues — and the few places that want a
selection back out of the text decode it there. Records, the search index and
a favorite's stored selection all hold the text, which is also what the
thumbnail url carries and what the R2 key hashes.

Also collapses the two identical encoders into encodeConfiguration, and
replaces the hand-rolled cyrb53 with crypto.subtle: Workers implements
WebCrypto, so a truncated SHA-256 needs no dependency. That makes the key
async, which every call site already was.

Stored records only carry the new shape after a library reload. Favorites are
converted in place by 0008, which json_each walks in parameter order.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

Storing only the canonical form lost data: canonicalizing drops every value
that equals a parameter's default and every hidden one, so a favorite could
not carry a string parameter the user typed. The row keeps the selection as
made again, and 0008 is dropped rather than reversed — it never ran anywhere.

The canonical form the thumbnail needs is computed in getFavorites instead,
against the insertable's current parameters. Storing it alongside would go
stale the moment a reload changed what a parameter defaults to, and computing
it per card would be a parameters query per row.

Also unnests the card's part name and number, which rendered a div of p's
inside the p the title is, and so tripped a hydration warning. It is a line
under the title, so it belongs beside it in the stack.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

A configuration only ever had to name one path segment, and url-encoding it
does that — the way Onshape spells one — so the hash bought nothing but a
digest to explain. Encoding keeps `/` and `;` out of the path, and a key stays
readable, which a truncated digest never was.

Drops toConfigurationKey and DEFAULT_CONFIGURATION_KEY, and with them the only
reason thumbnailKey's callers were async. Also notes on canonicalizing that it
is lossy, which is why a stored selection is never only its canonical form.

Thumbnails stored under the old hashed keys are orphaned: nothing reads them,
and the configurations that want them re-render on demand.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

Every comment over two lines is cut to two without losing its reason, the
padded JSDoc blocks collapse to one line, and the comments that narrated the
line under them or restated a component's own name are gone.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

Inline style was re-stating what Mantine already exposes: flex, miw, maw and
display are style props on every component, Center is a grid that places its
one child, and Box takes miw where a bare div needed style to.

What is left is genuinely custom, so it is named once rather than written out
at each use: NO_SHRINK for holding an icon beside text that can outgrow its
row, and maskedImage for painting an asset in the current text color.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

A component's props were declared three ways — an interface, an object literal
at the signature, a destructure at the signature — so a reader had to find out
which before reading the props. They are all interfaces now.

Defaults were split the same way: destructure the rest of the props, then a
`props.x ?? fallback` line below for the optional ones, which reads as though
it were doing something more than defaulting. They land in the destructure.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

One file held 34 declarations and every layer of the build card at once, while
every other feature here is already several files. It becomes five, each one
thing the card is made of: the badges and hover card that open it, the issue
severities it summarizes, the admin toggles, the parsed metadata, and the two
row primitives all of them share.

The seams are one-way — sections is a leaf, issues and parsed sit above it,
admin above those, and the card composes them — so the imports say which layer
a component belongs to.

Every component here took its props as a literal type at the signature; they
are interfaces now, like the rest of the frontend.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL

The big components were doing three jobs at once: holding state, wiring
mutations, and drawing. What was state or wiring is now a named hook —
useInsertSelection, useSetDefaultConfigurationMutation, useHomeSections,
useDefaultConfiguration, useReportSelection — and what was a self-contained
block of markup is a component: InsertMenuFooter, SectionAccordion.

useMenuTitle replaces the same updateModal effect written out in two menus. It
takes an undefined name to mean not yet known, which is what the effects it
replaces guarded for: the opener has already set a real title, so writing an
empty one over it would blank the header until the query lands.

The last components taking an object type literal have props interfaces too.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018tgfuCHjVFvejjMe7WmEFL