#491 rewrote smoke goldens so they stop pinning the tool version, and landed
on main between the last sync of #398 and its merge. The two touched different
files, so they merged cleanly, but the catalog's claims pin the checksum of the
golden that backs each of them, and fifteen of those now named a file that had
changed underneath them.

The first `Release prerequisites` run on main caught it: `catalog-valid` failed
on `finding-baseline-workflow.golden.json`. That is the gate working as
designed, and it would have refused the next tag.

Refreshed with `catalog-validate --refresh`; no claim changed, only the
recorded checksums.

Co-Authored-By: Claude Fable 5.1 <[email protected]>

The catalog stored a SHA-256 for every fixture and expected-result file a
claim names. Those files are regenerated routinely, so every golden update
left the catalog stale, and the `catalog-valid` gate then refused the next
release until someone rewrote the checksums.

The previous commit on this branch did that rewrite. This one removes the
reason for it.

A file in this repository is already content-addressed by git: a release tag
and a path identify exact bytes, and the report records the commit. The stored
checksum was a second, hand-kept copy of what git knows. It could not catch
anything git does not, and the only thing it could do was fall behind.

It could also not be kept in step by checking earlier. `make test` already
verified the checksums on every pull request. #491 changed goldens and #398
added the catalog; each passed CI on its own, they touched different files so
they merged cleanly, and `main` failed the gate. A refresh step in `Update
Smoke Goldens` only helps when that workflow is what changes a golden.

So the catalog names files by path only, `catalog-validate --refresh` and its
step in the goldens workflow are gone, and that workflow no longer needs Go.
What is checked is what can actually go wrong: a claim naming a file that was
renamed or deleted fails `make assurance-catalog` and
`TestRepositoryCatalogIsValid` in `make test`, and both refuse a run that
checked no files at all (ADR-0044).

Still pinned, because this repository's history cannot address them: a git
input by revision and a container by digest.

The catalog and report keep their schema versions. No report has been
published, the site does not read the removed field, and the catalog has no
reader outside this repository. ADR-0047 carries a dated amendment.

Co-Authored-By: Claude Fable 5.1 <[email protected]>

The first prerequisites run failed TestReportGoldens on Windows only. The
goldens are compared byte for byte, and a Windows checkout rewrites text
files to CRLF; converting a golden to CRLF locally reproduces the same
"golden differs" failure.

The assurance testdata is marked -text so it is checked out exactly as
committed on every platform. Scoped to this directory on purpose: the rest of
the repository's testdata has passed on Windows as it is, and changing how it
is checked out is not this change's business.

Co-Authored-By: Claude Fable 5.1 <[email protected]>

Nothing ran the unit suite on Windows except the portable stability workflow,
which was manual and last ran in July. Making it a release gate ran it again,
and three tests had gone red there in the meantime.

- SwiftPM published its evidence patterns backslash-separated on Windows. The
  candidate list was built with filepath.Join and is used both to read files
  and as the detector's evidence patterns, so the support matrix and `bomly
  plugin list` said different things depending on the machine. The list is
  slash-separated now; every read already joins an entry onto a directory with
  filepath.Join, which converts the separators for the host.
- The SBOM detector test looked for a file path inside an error that quotes
  it. Quoting doubles a Windows path's backslashes, so the bare path was not a
  substring there. It looks for the quoted path.
- The file-cache contract test asserted Unix permission bits. Windows has
  none: Go reports 0777 for every directory whatever mode it was created with.
  The mode assertions are Unix-only; the containment half of the contract is
  still checked everywhere.

Co-Authored-By: Claude Fable 5.1 <[email protected]>

Review pointed out that removing the per-file checksum from report claims
without raising the report's schema version contradicts the rule this
framework wrote down for itself: only an optional addition may keep a version.

The rule was stated without its starting point. A version names a shape
someone already holds, and for the report that begins with the first report
committed to docs/assurance/reports -- from then on bomly.dev has mirrored a
document of that shape. No report has been published, the site never read the
removed field, and its sync does not validate it, so there was no reader to
break. Raising the version would have made the first report ever published a
"v2" beside a "v1" that never existed.

The developer guide now says that in so many words, and the ADR amendment
records it. More usefully, TestPublishedReportsParse makes it mechanical:
every committed report is read with the current parser, which rejects unknown
fields, so once a report exists a change that strands it fails in `make test`.
Checked by mutation -- a report in the current shape passes, one carrying the
removed field fails, and an index naming a missing report fails.

Also marks the sentence in ADR-0047 that the amendment supersedes, so a reader
does not have to reconcile the two.

Co-Authored-By: Claude Fable 5.1 <[email protected]>

The guard added in the previous commit had two holes, both found in review.

It parsed only the reports the index lists, so a report file the index did not
mention was never read. And it relied on the strict parser alone, which
rejects a removed or renamed field but not a newly required one: such a field
decodes silently as its zero value, so an incompatible addition under the same
schema version passed.

Both close on one property. Every file in docs/assurance/reports must parse
under the current schema and encode back to exactly the bytes it was read
from. A removal or rename fails the parser. A newly required field appears in
the re-encoding and the bytes differ. An added optional field is omitted when
empty, so it passes -- which is precisely the one change allowed to keep a
version. The check asks the encoder instead of listing fields, so it cannot
fall behind the struct. The directory and the index must also agree in both
directions.

The same property is asserted on the fixture reports, so it is exercised
before the first report is published.

Checked by mutation: a current-shape report passes; a field added without
omitempty fails; the same field with omitempty passes; an unlisted report
fails, and is still read when its shape is broken; a listed report that is
missing fails.

Co-Authored-By: Claude Fable 5.1 <[email protected]>

This pull request fixes a stale-checksum failure. In its first review round a
reviewer noted that dropping a report field without raising the schema version
contradicted the written rule, so a guard was added to hold that rule
mechanically. Every round since has been about the guard:

- round two: it skipped report files the index did not list, and it let a
  newly required field through, because such a field decodes as a zero value;
- round three: it did not cover the index, and its byte comparison would fail
  on a Windows checkout, which rewrites line endings.

The last one is the reason to stop. That failure would land in the portable
suite, which gates a release, as soon as the first real report was committed.
A guard that byte-compares published data had become a release risk in its
own right, and it was being refined in place of the change it was attached to.

So it is removed, and the reasoning is recorded where the next person will
look: beside TestSchemaVersionsArePinned and in the developer guide. The rule
itself is unchanged and is still stated there, including when a version starts
to bind. It is held by the pinned-version test, by the report goldens, where
any change of shape shows up as a diff in review, and by the reviewer. A
mechanical guard, if wanted, is its own change, and should compare the
report's JSON schema against the one recorded at the first published release
rather than committed bytes.

Co-Authored-By: Claude Fable 5.1 <[email protected]>