feat(source): add bounded read-only mixed filesystem discovery

2026-09-24 21:57:16 -07:00
parent f83f3bf45b
commit 3fd87bd8be
8 changed files with 739 additions and 31 deletions
@@ -0,0 +1,48 @@
+# Mixed CLI core discovery implementation
+
+Base bab6f96a; selected root contract f810fb40. This contribution owns core only.
+source_files::load_mixed_source returns opaque MixedSourceSnapshot with immutable
+bundle/diagnostics views. It consumes a fresh private DiskProject after final
+checks and drops root/expectations/reachability; no reachable-bundle clone, model,
+index, save or recovery authority escapes. The byte allowance is clamped to the
+existing64MiB ceiling before disk reads; other caller fields are preserved.
+
+The existing discovery loop has a private borrowed ordinary/mixed selector.
+Ordinary overlays, aggregation, limits, final compilation and journal behavior
+are preserved. Mixed selects a shared Budget/ModelCounts phase and skips ordinary
+compilation. Syntax/budget failure records the original file diagnostic and stops
+parsing later siblings, while final disk/physical/journal checks still execute.
+Disk/encoding/path failures retain existing classifications and safeguards.
+
+Both producer modes start once at zero. Each inspected file reserves its phase
+slot before scanning. Bytes/tokens/references/numeric/syntax/model work accumulate
+across units. Remaining text/value/owner reservations remain conservative after
+unit destruction; in-parser temporary/UUID/key releases are unchanged. No edit
+positions or complete model are constructed during discovery. Final public
+semantic checking is a separate pass with the original caller policy.
+
+A narrow shared parse_unit helper centralizes exactly the existing Parser setup,
+unit(entry,true) call, state extraction and Parser drop. Complete mixed parsing
+keeps all subsequent metadata/projection/traversal/accounting operations. Source
+review finds no semantic blocker; helper-frame effects still require parent's
+actual existing stack gates and prologue comparison, not a source-only claim.
+
+Three internal discovery tests cover exact132-byte/31-token/3-event chain totals,
+conservative text32/metadata3 after unit drops, all temporary balances, original
+policy capture, cumulative6/5 entry boundary with no later-file remapping, syntax
+failure terminality and positive cumulative numeric/reference work. The tight
+entry profile is explicitly a fresh internal fixture, not a public policy knob.
+
+Five external mixed_source_files tests exercise real filesystem capture and
+independently authored complete model equality (nested ordinary participants,
+cross-file calculations and scoped parameters), original bytes, four missing-ID
+classes, malformed syntax, unparsed later sibling plus preserved pending journal,
+missing/escaping/cyclic includes, file/byte/depth/edge boundaries, invalid UTF-8,
+nonregular files, case collisions and hard-link aliases. CLI worker separately
+owns OA/SA rich command fixtures/report/actual executable tests and documentation.
+No old test body or expected outcome was changed.
+
+Only owned Rust files were formatted directly with skip_children=true. No Cargo,
+Rust compiler, rustdoc, CLI execution or remote operation ran in this contribution.
+Parent owns all runtime, production ordinary-regression, stack, lint and combined
+source acceptance. Existing depth/platform/accounting limitations remain intact.