[OmarAlJarrah]

Summary

A path item lowers to no node of its own, so what it declares that its operations have no home for
(its summary and description, its x- extensions and its undeclared keys) is kept under Unmodeled
on every operation it mounts. This changes how those entries are built and how the documentation
pair is keyed.

Built once per mount. mountPathItem computed the entries inside its per-operation loop, once
per operation and per layer, though none of them depends on the operation. Each layer's keys were
censused, sorted, looked up, rendered from YAML to JSON and reported again. With a counter at each
step, a path item writing two undeclared keys over eight operations ran the census 8 times and built
16 findings, of which compile.Diags kept 2. A $ref'd item with a sibling key of its own (two
layers) ran applyPathItem 16 times and built 24 findings, keeping 3. applyLayers now runs once
per mount into one map, beside itemServers and the precedence check, which were already once per
mount. keepOn then copies each entry onto each operation, cloning its value so no two operations
share bytes. The note that an operation holds its path's documentation stays per operation, since
it is sited there.
applyPathItemDocs no longer makes that note, so the carrier's provenance field had no reader left
and is removed, along with onNearestNode's context parameter.

Documentation keyed beneath the item's scope. The pair was kept under the fixed keys
openapi:pathItemSummary and openapi:pathItemDescription. An operation's census spells the same
keys for an undeclared key of that name, and the path item's text, written after the census, replaced
it. The operation's value reached the IR in no form while its warning said it was kept (#1323). The
pair is now keyed beneath the carrier's scope, as the item's extensions and undeclared keys already
are: openapi:pathItem/summary and openapi:pathItem/description. The census escapes a / in the
key it keeps, so no key of the operation's own can spell one. An item mounting no operation already
keeps its pair beneath a scope holding its own pointer (openapi:pathItem/<item pointer>/summary),
so applyPathItemDocs now keys every carrier by annotation.EntryKey(into.scope, keyword), and the
carrier's operation flag, which chose between the two spellings and the note, is removed. The
path-item-docs rule says so and now lists its tests.

No entry replaced. Building once moved the census's check for an occupied entry off the
operation's own map, onto the map the mount's entries are built in. keepOn now copies through a
new annotation.CopyEntriesInto, which never replaces an entry the carrier holds. It reports a
refused copy under openapi/unknown-key-entry-taken, in key order as the census does, but with a
message of its own: the operation keeps no copy, where the census's message says the IR holds none,
which would be false while the mount's other operations hold it. It clones each value it writes.
The scoped keys mean no input reaches that branch today. It is there so a later key that clashes
is reported rather than silently replacing the operation's entry. The note about the path's
documentation is made only when that entry is the one the operation holds.

What a consumer sees:

• Two Unmodeled keys on operations are renamed, as above, so IRVersion moves to 0.30.0 (see
  Breaking).
• A path item's own findings are reported before its operations' rather than after the first
  operation's. That reorders the diagnostics in two conformance goldens.
• Nothing else moves. Besides the rename, the goldens' only change is that order.

Closes #414.
Closes #1323.

Breaking

IRVersion moves from 0.29.0 to 0.30.0. The JSON shape is unchanged, so no pin asks for the bump,
but a consumer built against 0.29.0 that reads openapi:pathItemSummary or
openapi:pathItemDescription on an operation finds the path's text gone, silently: the failure
docs/ir/ir-version.md says a version exists to report. 0.26.0 was bumped for the same reason,
when Unmodeled entries a consumer read stopped being written. Where an operation writes an
undeclared key of one of those names, the old key now holds that key's own value, which 0.29.0 let
the path's text replace. The generations row repeats 0.29.0's shape and names and states the
meaning, and docs/ir/version-history.md says what a consumer pinned to 0.29.0 sees.

Test plan

• TestPathItem_KeepsWhatItDeclaresOncePerMount reads the walk's diagnostics before the dedupe,
  through a new serviceAsReported helper that shares walkService with serviceWithGrouping.
  Over an inline item with three operations writing a summary, a $ref'd two-layer item with two
  operations writing a summary at its use site and a description in its referent, and an item
  writing only a description, it requires one openapi/unknown-object-key per source key, one
  documentation note per operation, every entry on every operation, an operation's own entry beside
  them, and each operation owning its bytes.
  • On main (7f84912) it fails on the once-per-key assertion, on the note, which main makes
    twice on each operation of the two-layer item, and on the renamed keys.
  • Planted at 7adac1b, each of these fails it: the clone dropped, the note dropped, the note made
    for each keyword that landed, only summary checked, only description checked, entries
    written onto the first operation alone, the operation's own map replaced, the entries rebuilt
    for each operation, and the fixed keys restored.
• TestPathItem_DocsKeepApartFromAnOperationsOwnKeys gives an operation undeclared keys spelled
  pathItemSummary and pathItem/summary under a path with a summary, and requires all three
  values under keys of their own and no openapi/unknown-key-entry-taken anywhere. On main it
  fails on the operation's pathItemSummary holding the path's text. Copying the entries onto an
  operation twice fails it, which an assertion at the operation's own key alone would let pass,
  since the refusal is sited at the path item's.
• TestCopyEntriesInto_KeepsWhatTheCarrierHolds and TestCopyEntriesInto_ReportsInKeyOrder hold
  the copy to keeping the carrier's entry, reporting the refused one at its own pointer, naming the
  holder and claiming no loss from the IR, cloning, and sorting 64 refusals by key.
  TestKeepOn_AnnouncesOnlyDocumentationThatLanded holds the note to the entry that landed, a
  description beside a refused summary included, and to its code. Each of these mutants fails one
  of them: the held entry overwritten, the refusal left unreported, the refusal reported with the
  census's message, the sort dropped, the clone dropped, the note made whatever landed, the note
  given up at a refused summary, and the note made under another code.
  • The key-order test refuses 64 keys written in reverse. With the sort dropped it failed 400 runs
    of 400. With eight keys written in order, as it first had, it passed 46 of 400: Go iterates a map
    of eight or fewer entries as a rotation of the order they were written in.
• TestUnknownKeys_PathItemDeclaredFieldsAreNotUndeclared refused every key beneath
  openapi:pathItem/ as a census entry for a declared field. It now admits the documentation pair
  alone, and still fails when the census is planted to treat query or servers as undeclared.
  Planting the same for summary passes it, since the census then finds the docs entry at its own
  pointer and stays silent; TestPathItemFields_MatchTheLibraryModel fails on that plant.
• TestApplyPathItemDocs_WithoutRootNode now runs through keepOn as well, so it still holds that
  nothing is announced when nothing was kept.
• morphic compile with a main build and a build of this branch over 2,068 specs: the 227 under
  testdata/, 1,840 real-world OpenAPI 3 specs, and the Cloudflare API schema @319dc756. It ran at
  main b206dc9, at 7f84912, which bumps the YAML library, at be889c6, which reworks extension
  promotion and adds two of the specs, at 02c972b, which retypes the local time formats, and at
  117d29e, which keeps an unmounted item's documentation, each after merging it; the last three
  runs also follow the bump. With the two keys renamed in main's
  output, and its irVersion stamp read as the branch's where the branch bumps it, exit codes, the
  IR and the multiset of diagnostics are identical on every one. Ten of the specs hold the renamed
  keys, five of them real-world.
  Diagnostic order differs on three, each a path item with undeclared keys: the two goldens above
  and testdata/openapi/unknown_keys.yaml, which has no golden.
• The golden diffs hold nothing else. After the first commit, each changed golden equals main's
  but for the order of its diagnostics. After the second, renaming the two keys in the first
  commit's goldens leaves no difference.
• TestIRVersion_PinsTheEncodedShape and TestIRVersion_PinsTheNames pass with the 0.30.0 row,
  whose hashes equal 0.29.0's, because it states its meaning.
• The gate runs in CI.