[OmarAlJarrah]

Summary

A path item that mounts no operation, whether under paths, webhooks or a callback expression,
already keeps its servers, extensions and undeclared keys. They go on the nearest node that holds an
Unmodeled map: the service, or the parent operation's HTTP binding for a callback. Its summary,
description and parameters reached the IR in no form, though. An item that wrote only those drew
no diagnostic at all:

paths:
  /p:
    summary: PATH SUMMARY
    description: PATH DESC
    parameters:
      - {name: trace, in: header, schema: {type: string}}

On main, morphic compile exits 0 with an empty stderr, and none of the three appears in the
document.

Summary and description. They were dropped on purpose. The fixed keys an operation holds them
under (openapi:pathItemSummary, openapi:pathItemDescription) would collide when two such items
share one carrier. They now take the item's scoped key, the way its extensions do:
openapi:pathItem/<item pointer>/summary and .../description, with reason no_ir_home. The
warning that announces the item covers them, so they draw no info finding of their own.

Precedence across the layers of a $ref'd item now counts them whether or not the mount lowers an
operation. So a summary written on both sides of a $ref is reported as an
openapi/path-item-ref-collision, and the use site's summary is the one kept.

Parameters. Item-level parameters lower only by merging into each operation the item mounts, so
an item with no operation kept nothing of them. Each entry is now kept as written, a $ref and the
fields beside it included. The key is openapi:pathItem/<item pointer>/parameters/<name>/<in>,
which is the path its parameter ID would be derived from (ids.ParamPath, which ids.Param now
uses too), and a repeat is numbered the same way its ID would be. Per-entry keys are needed because
one mount's list can come from several layers, and the nearer layer wins slot by slot.
paramLedger's counting is split out of next so both paths share it. That also reports a repeat
in a list another document writes, as the operation path already does. An entry whose $ref
resolves to nothing has no name or location to key it by. The resolver reports it, and it is kept
nowhere, as on an operation.

Rules and corpus. The path-item-docs, path-item-servers and parameters rules in
docs/compilers/openapi.md now say what an item with no operation keeps, and path-item-docs
gains a Tests line. path-item-docs.yaml gains an item with no operation, and param-identity.yaml
gains a component item with no operation that two paths mount. This way the corpus sweep reaches
both. Its order check compares their findings but reads no node's Unmodeled, so the kept entries
rest on their keys instead. Each key holds its item's pointer, the unit tests assert each item's
entries apart, and the docs test holds every key on a carrier beneath the scope of an item that
carrier holds. A key two items shared would fail in either order. DuplicateParameter's doc
comment, paramLedger's and the parameters rule each list where a repeat is checked, and now
name the list such an item keeps.

Closes #383
Closes #1080

Left as they were, each stated where the next reader meets it:

• Whether a path's documentation should get a typed home in the IR, rather than Unmodeled, is
  ir: decide whether a path item's summary and description get a typed home #1319. The path-item-docs Status and applyPathItemDocs now cite it, in place of openapi: a path item mounting no operation drops its summary and description in silence #383.
• A repeat in a list that only a $ref reaches is still unreported (openapi: a list only a $ref reaches is never checked for repeats #880), for an item with no
  operation just as for one with an operation. The parameters rule's Status already says so.
• The warning for a layer another document declares is still sited in the source document (openapi: a path item mounting no operation is reported in the referenced document at the mount's pointer #1164),
  as preserveUnmountedPathItem says.

While sweeping the fields a path item can write, I found that a null additionalOperations entry
makes the whole document undecodable, on main as here. It is a decoding defect apart from this
change, and is filed as #1324.

This overlaps #1327, which keys an operation's pair beneath its scope too
(openapi:pathItem/summary) and builds a mount's entries once. The two edit the same lines of
pathitem.go, pathitemref.go and the path-item-docs rule, so whichever lands second resolves
them. The resolution can key every carrier by annotation.EntryKey(into.scope, keyword), and drop
the carrier's operation field, which only chooses that key and the operation's info finding.

Test plan

• New tests in operation:

  • TestPathItem_WithNoOperationKeepsItsDocs and TestPathItem_WithNoOperationKeepsItsParameters
    cover each of the three routes. Each route has its own text, and each entry is checked for key,
    value, reason and provenance. They also check that a $ref entry is kept as written and that a
    name holding / and ~ is escaped as one segment. The docs test holds every key on each
    carrier beneath the scope of an item that carrier holds, so a key two items share fails under
    any spelling.
  • TestPathItem_WithNoOperationNumbersARepeatedParameter
  • TestPathItem_WithNoOperationReportsARepeatInAnotherDocument
  • TestPathItem_WithNoOperationKeepsNoUnresolvedParameter
  • TestPathItem_WithNoOperationKeepsEachMountApart

• Updated tests:

  • TestPathItem_WithNoOperationAndNothingToKeepIsSilent had pinned a docs-only item as silent.
    It now holds an empty pair to that instead.
  • TestPathItemRefSiblings_UnmountedCollisionKeepsTheUseSite had pinned "neither summary is
    kept". It now holds the use site's summary and parameter, and the referent's description and
    other parameter, each at its own pointer, with each collision reported.
  • TestPathItem_UnmountedAnnouncementSaysWhatLanded gains a docs-only row and a parameter-only
    row.

• Twelve mutants were planted, one at a time, each against go test ./compilers/openapi/internal/operation ./compilers/openapi/e2e -run 'TestPathItem|TestParams|TestConformance|TestGolden', and every one
  reddened at least one named test:

  • dropping the keepItemParams call
  • dropping the docs again for an item with no operation
  • using the fixed doc keys on every carrier
  • emitting the info finding for an item with no operation
  • ignoring precedence for parameters
  • not numbering repeats
  • keeping the resolved node instead of the written entry
  • siting provenance at the mount
  • discarding the repeat finding
  • keeping an unresolved entry
  • taking the docs out of precedence
  • leaving the name unescaped

  The kept-entry checks compare an entry's provenance whole. Two more plants each set a Node on
  the provenance of a kept parameter or a kept summary, and each reddened the tests that cover it.
  A plain pointer check would have missed both.

  Three more plants each wrote, beside the scoped pair, a key no item's pointer scopes:
  openapi:pathItemSummary, openapi:pathItem/summary, and the latter on a callback's carrier
  alone. Each reddened TestPathItem_WithNoOperationKeepsItsDocs. The check the scope check
  replaced named the first spelling alone, and passed the second.

• go test ./compilers/openapi/e2e -run 'TestGolden|TestConformance' -update changed only the two
  extended cases. Their diffs hold the new entries and one warning per item with no operation.

• The spec above, compiled with morphic compile --verify, now holds
  openapi:pathItem/paths/~1p/summary, .../description and .../parameters/trace/header, and
  draws the one warning. The same item under webhooks, and as a callback expression, lands on the
  service and on the parent operation's binding respectively.

• A spec with seven items that mount no operation compiled the same with every declaration reversed:
  identical kept entries and the same findings. The items sit under paths, webhooks and two
  callbacks, and include a $ref with fields beside it and a component two paths mount.

• make gate is left to CI.