v0.2 Revision Candidates

Adjudicated 2026-09-01 — the cycle is complete and v0.2 is declared (see ../REVISIONS.md). Outcomes: A1 → v0.1.4 (D31) · A2 → v0.1.4 clarification · A3 → v0.1.4 registry (E1xxx/E2xxx) · B1+B2 → v0.1.5 (D32, E4093 now static) · B3+B4 → v0.1.5 clarifications · C1 → v0.1.8 (D33, grammar + wasm updated) · C2 → v0.1.7 (D29 amended) · C3 → v0.1.6 (split/contains/ ends_with) · C4 → no language change (idiom stands) · D1/D2/D4 → implemented (E4090/E4093 static checks, matches evaluation + static E4071) · D3 → remains a documented linker limitation.

Findings collected while implementing Phases 2–4 against the frozen v0.1 specification and validating it on real-world corpora in Phase 5. Nothing here changes v0.1: each entry is a candidate for the v0.2 revision cycle, to be adjudicated one by one through the standard process (charter decision + affected chapters + REVISIONS.md in one change). Entries are ordered by category, then by weight of evidence.

A. Spec defects — the frozen text conflicts with its own corpus

• A1. Strict S ⊑ T static assignability rejects the guide and benchmarks (§3.18, §4.4). port: 9000 + i (with i: 0..<3) infers int, which is not ⊑ 1..65535; a string template is not ⊑ a pattern type. Both forms appear in normative examples. The reference implementation keeps the corpus sound with two precision devices — interval arithmetic over int ranges for +/-/*, and deferring same-kind refinement targets (pattern / range / literal set) to binding when membership is statically unprovable. v0.2 should standardize both (or an equivalent rule) in §3.18/§4.4.
• A2. Unit-literal lexing overlapped numeric tokens (§2.7 vs ch. 11). 0o755, 1e3, 1e-3 lexed as unit literals under the published grammar, contradicting §2.7 (“only decimal forms take units”, float exponents are part of the number). The tree-sitter grammar was fixed to §2.7 by construction (no lookahead); ch. 11’s EBNF should state the same token-boundary rule explicitly.
• A3. No parsing-error code band in §12. The registry starts at E3xxx; tools need a code for “syntax error” (the implementation uses a provisional code). Reserve and define an E1xxx band.

B. Underspecified semantics surfaced by use

• B1. Module-const records shared into several roots (§7.5). A const base: Topology embedded in two outputs gives its internal refs const-rooted places — E4093 territory, but nothing defines what embedding a const-rooted record value into an output means for the references it carries. The working idiom is a constructor func returning the literal (each output binds its own copy). v0.2 should either bless that idiom normatively and add the static E4093 check, or define re-rooting semantics for embedded values.
• B2. Reshaping unbound literals (§4.12, §4.8). A constructor func’s result is an unevaluated literal until it reaches a typed position; chaining with / comprehensions over it before binding has no defined semantics (the implementation supports plain with on such values, nothing deeper). Either define the lazy-structure semantics or state the restriction.
• B3. Declaring-module scope for member expressions (§8.3). Imported types’ derived members and asserts must evaluate in the scope of the module that declared them (sibling consts, namespaces, $referrers targets) — the implementation enforces this; the chapter implies but never states it.
• B4. Continuation lines in assert (§2.9, ch. 11). After assert name: a newline continuation may not begin with (; the working form wraps the whole condition in parentheses opened on the colon line. Specify the continuation points precisely (the grammar has them; the prose does not).

C. Extension demands from the real-world corpora

• C1. Expression-position access to keyword-named members (§4.3). Real documents use type as a member name. Declaration positions accept it; expressions must fall back to $this["type"]. Consider contextual keywords for member access (x.type).
• C2. Numeric leniency at the interchange boundary (§10). Real documents serialize whole floats as integers ("frequency": 500). Schemas work around it with int | float. Consider: a float-typed binding position accepts an integer lexeme exactly representable in binary64.
• C3. std.string growth (§13.6). Validating edge keys of the form "a->b.c" needs split (and would use contains/ends_with). One-hop operations stayed sufficient otherwise — std.graph remains unjustified (OQ3 verdict unchanged).
• C4. Empty-string sentinels as an idiom. Real parameter bags use "" for “not configured” where a name is otherwise required; the "" | Ident union reads well and needs no language change — worth a guide/idiom note.

D. Implementation gaps to close (no spec change needed)

• D1. E4090 (embedding-site context bounds, §7.3) has no static check yet.
• D2. E4093 (ref navigating a module const) has no static check yet (see B1).
• D3. Generic type arguments that reference importer-local types resolve in the exporting module’s scope and can miss (documented limitation in the module linker).
• D4. matches evaluates only where the checker types it; the engine lacks a matches case.

Evidence base

Phases 2–4 test suites (251 checks across eight runners) plus the Phase 5 sweeps: the three committed domain examples under examples/ (service graph, fixture generation, synthetic network fabric with scale and corruption probes), and a local-only schema additionally validated against the full proprietary fixture corpus — 178 documents including the complete real set (artifacts kept out of the repository by security policy).