feat(file): identity is the rel — the mint flips, and the trace records the binding - #596

…rds the binding

Phase 2, PR 1 of the resource-construction plan (#584). File identity mints from SourcePath.Rel() in the
ruled opaque form: "file:" + rel, slash-canonical on every platform (4.1's amended file-scheme row and
sketch, landed in phase 0). The fsroot is a run parameter, so the rel half is what serializes and the same
graph relocates across prefixes; no volume, home layout, or separator variance enters identity. The
native form is access, not identity -- consumers reach it through SourcePath.Abs(), never by parsing the
URI (the #547 rule). Escaping rels ("../...") still mint until PR 2's plan-space grammar refuses authoring
them.

The 2026-08-18 revert's consumer hunt resolved empirically to ONE seam, not four: writ's readback parsed
the URI's embedded machine-absolute path and keyed drift attribution by it. The design-true fix: the
trace's ledger snapshot gains Root -- the run's bound fsroot, stamped by the executor at capture (both
capture sites route through the extracted captureLedgerSnapshot) -- and recordedIdentity derives its
native keys by joining the recorded root with each entry's rel payload. Section 5.5 records the doctrine:
the graph stays root-free (intent); the trace records the binding (observation), because a recorded rel is
interpretable only against the root that ran it.

The foreign-scheme pin re-asserts against the rel form (file:https:/example.com/x), which is
platform-neutral -- this is one of the three Windows baseline failures and is expected to clear. The plan
carries the phase-2 review findings (failure diagnosis from the live CI logs, the grammar/harness
co-landing constraint, the three-PR slicing) and this PR's record.

Verified with make check (103 packages ok, 0 failures) and make vet under GOOS windows. Windows baseline
expectation: 3 -> 2 (the remaining two are path-as-text at their authoring/narration surfaces, PR 2's
work).

Refs #584, #547, #546.

…led form

The Windows re-diagnosis of PR 1's first CI run: the expected 3 -> 2 came back as 3 with a NEW name --
TestDiscoverRegular_PathAndOwnSpecific_SameIdentity panicked in fsroot.makePath, because the test hardcoded
the retired emitted form ("file://" + abs). Stripping the new "file:" prefix from that spelling leaves
"//C:\..." on Windows, which reads as UNC and cannot be made root-relative. On unix the same spelling
survives by accident (the extra slashes clean away), which is why only Windows caught it.

The pin now asks the resource for its own specific (ReachabilityURI()) instead of spelling one, so the
round trip follows the mint permanently. The foreign-scheme pin cleared as predicted, so the Windows
expectation for this branch remains 3 -> 2.

The plan also records the run-from-elsewhere caveat (USER ruling, 2026-08-21): the existence check reads
the CONSTRUCTION environment's root and feeds the construction-time absolute, so a graph saved and run
later from elsewhere verifies against the wrong root -- plan-and-run-in-one-session flows merely hide it.
PR 3's rel-first activation binding at the run boundary is therefore load-bearing for relocation, not
closure hygiene.

Verified with make check (103 packages ok, 0 failures) and gofmt clean.

Refs #584, #547.

The second Windows re-diagnosis of PR 1: the foreign-scheme identity pin failed with
file:./https:/example.com/x -- Windows filepath.Clean guards a first segment containing a colon with a
leading "./" (the CVE-2022-41722 drive-relative disambiguation), and the guard rode through ToSlash into
the rel, so the same input minted different identities per platform. Exactly the class #547 predicted
("the sixth is waiting"), and exactly where its rule says to fix it: in the type, not at a call site.

fsroot.canonicalRel (slash-path Clean after separator normalization) now renders every rel -- both
makePath branches and fsroot.NewPath -- so OS guards, duplicate slashes, and dot segments alike collapse
to one neutral form. A platform-neutrality pin (TestRoot_NewPath_ColonSegmentRelIsPlatformNeutral) runs
the colon-segment case across every root variant on all three CI platforms. Recorded on #547 as the
sixth-discovery table row.

The plan carries the sixth-discovery record and PR 2's new obligation: the plan-space grammar reserves
the root-qualification spelling for #597 (the named multi-root design question, filed under #546), so the
little language does not break when (root-name, rel) identity lands.

Verified with make check (103 packages ok, 0 failures), make vet under GOOS windows, gofmt clean. Windows
expectation for this branch: 3 -> 2.

Refs #584, #547, #597.