Phase 1 / Etch / Cross-file import (resolver pass-1)

[guysenpai]

M1.0.7 — Cross-file import (resolver pass-1)

Brief : briefs/M1.0.7-cross-file-import.md (Status: CLOSED).

Implements cross-file import in Etch end-to-end: graduates import parser-up (lexer → AST → parser), builds the module dependency graph + topological order + cycle detection (E0108), a per-module byte-keyed exports index + import binding (E0103/E0104/E0107), and cross-arena component resolution that unblocks the E1793 false positive for .prefab.etch.

Execution (E1→E6, one commit per slice)

• E1 feat(etch): graduate import to kw_import token — kw_import out of non_s3_keywords.
• E2 feat(etch): add ImportDecl AST node + arena slabs — ImportDecl/ImportItem + slabs + addImportDecl.
• E3 feat(etch): parse import directive, 4 forms + recovery — parseImportDecl/parseImportItem, dispatch, sync-set.
• E4 feat(etch): import module graph, topo order + E0108 cycle — deriveModulePath, graph, iterative DFS.
• E5 feat(etch): per-module exports index + import binding — ProjectContext exports, bindImports (E0103/E0104/E0107).
• E6 feat(etch): cross-file import resolution + E1793 unblock — imported TYPE_IDENT resolution + checkComponentInstance cross-arena.
• Docs: docs(claude-md): update for M1.0.7, docs: reframe foreign-field-type note as moot (not debt), docs(brief): close M1.0.7.

Closing notes

• What worked: import was the last reserved top-level keyword in non_s3_keywords; the table-driven lexer made E1 trivial. The M0.9 byte-keyed ProjectContext extended naturally to a per-module exports index; keying cross-file resolution on string BYTES (not per-arena StringIds) is the load-bearing idea and held end-to-end. The iterative DFS gives both the deps-first topological order and the E0108 back-edge in one pass. The cross-arena check localized to checkComponentInstance + one checkInstanceFieldForeign helper (every other validate pass already skips unknown item kinds; pass1Collect ends else => {}).
• What deviated from the original spec: one accepted deviation (Claude.ai round-trip, GO E5→E6) — test "selective import resolves a cross-file const" dropped (const is M1.0.8, not parseable here); cross-file resolution proven by the imported-component type test + the prefab unblock. FROZEN Acceptance edit logged under Accepted deviations (commit 6d4f344).
• What to flag explicitly in review: (1) cross-arena field-TYPE check (E1795) is complete for components — validateFieldsInDecl(.component_like) admits only builtin-POD field types, so the named-type branch is unreachable (forward-compat headroom, not a gap). (2) deriveModulePath strips optional src/ + .etch/typed-compound ext, /→.; handles bare fixture names and real src/ paths. (3) D-F (m.Type) and D-G (E0107) are wired-but-dormant, recorded so M1.0.8+ is additive.
• Final measures: no new benchmark (graph + topo + exports are O(files+imports+exports), off any hot path). validateProject cost did not move materially vs M1.0.6.
• Residual risk / tech debt left deliberately: none for the cross-arena field-type point (moot). Deferred-by-design (not debt): module-alias qualified m.Type (D-F), E0107 activation (D-G, M1.0.8), cross-file const import + test (M1.0.8).

Validation (Étape 4)

• All Scope deliverables (E1–E6) present
• No out-of-scope drift (import is the only graduated keyword; const/private/test/override untouched)
• Acceptance tests green in Debug and ReleaseSafe (import_parse_test, import_resolve_test, crossfile_prefab_import_test)
• Observable behaviour demonstrable (two-file project validates with zero diagnostics; .prefab.etch importing its components validates clean — asserted by the tests)
• zig build, zig build test, zig fmt --check, zig build lint green
• Closing notes filled; Status: CLOSED, close date set
• commit-msg + pre-push hooks green on every commit

🤖 Generated with Claude Code