Release/v1.2.0

[ikelaiah]

PR: v1.2.0 — Trust statement boundaries

Summary

This PR implements the v1.2.0 — Trust statement boundaries milestone from
ROADMAP.md. It ships Workstream A (statement-boundary hardening) in full,
plus the E boundary diagnostics (dialect_ambiguous, bracket/block balance)
with correct document/region scope. No existing golden changed: the fixture
corpus grew and every prior case remains structurally identical.

What's included

Statement boundaries (src/dialects.ts)

• Semicolons remain authoritative boundaries when present.
• Omitted semicolons are now split by dialect-aware statement grammar and
  control keywords; a newline is no longer required. Previously
  newStatementHere demanded a newline, so one-line semicolon-free statements
  such as SET @x = 1 PRINT @x SELECT @x collapsed into a single statement.
  They now parse into three statements.
• The SQLite SELECT RAISE(...) idiom stays one statement (RAISE does not
  split when it directly follows SELECT).
• The pre-existing INSERT…SELECT, UPDATE…OUTPUT, MERGE, and
  CASE-expression continuation guards are preserved.

Number lexing (src/tokenizer.ts)

• 0x / 0X hexadecimal literals (0x1F, 0x1_000).
• PostgreSQL-style underscore digit separators (1_000, 1_000_000).
• Leading-point fractions (.5) and trailing-point forms (1.) besides the
  existing 1.5 exponent forms.
• Dotted object names (dbo.t) and .. ranges are unaffected.

Diagnostics (src/ir.ts, src/dialects.ts)

• New dialect_ambiguous document-scoped warning on low-confidence detection
  ties (auto mode, several dialects tied). It carries no artificial span.
• New unexpected_end region-scoped warning when a block-terminating keyword
  (END, ELSE, …, UNTIL) has no matching opener, reported alongside
  unconsumed_input.
• Existing bracket/paren balance, unterminated_identifier, and missing_end
  diagnostics are asserted with region scope and valid spans.

Detection guardrail (src/dialects.ts, src/types.d.ts)

• detectDialect now reports tied (top scores equal), which drives the
  dialect_ambiguous guardrail.

Token-attribution fix (src/ir.ts)

• Attribution now marks by token identity instead of raw span. The previous
  span-based marking could hide filtered semicolons inside an unresolved
  region, so attribution did not always account for every body token. The
  fuzz suite now asserts the accounting invariant: resolved + ignored +
  unresolved + opaque always equals the body-token total, every edge has a
  recognised kind, and every node carries provenance.

Fixtures (tests/tests.ts, tests/boundary.ts, tests/fuzz.ts)

• New tests/boundary.ts with:
  • PROCFLOW_RANGE_FIXTURES — exact statement source ranges for terminated
    and valid unterminated statements across all four dialects, including the
    one-line semicolon-free cases;
  • new golden PROCFLOW_FIXTURES for CREATE VIEW … WITH, CTE-starting view
    bodies, ALTER TRIGGER, and semicolon-free one-line T-SQL.
• tests/tests.ts gained range-running plus direct assertions for number
  lexing, dialect_ambiguous, bracket/block balance diagnostics, and the
  hardened findBody headers (CREATE VIEW … WITH, ALTER object headers,
  multi-object scripts without GO).
• tests/fuzz.ts added the token-attribution, edge-kind, and provenance
  invariants.

Verification

• npm run typecheck — passes
• npm run build — passes
• npm run test:file — passes
• Golden suite — 156/156
• Fuzz suite — 400 deterministic mutation cases pass
• UI suite — 13/13
• Corpus metrics on the expanded fixture set: tail-unconsumed rate 0.00 %;
  unattributed-token rate 4.95 % (down from 5.05 % on the prior corpus);
  zero unknown-node fallbacks.

Deferred (per roadmap)

• Procedural parsing completeness (Workstream B)
• Query lineage accuracy (Workstream C)
• Data flow and dependency accuracy (Workstream D)
• Confidence re-score and layout replacement (v1.6.0 / v1.7.0)
• Catalogue, columns, and RDL (v1.9.0+)