fix(schema)!: span.bytes are UTF-8 byte offsets, every node and level (#179)

[rahlk]

Closes #179. Spec: docs/design/specs/span-bytes-are-bytes.md (D1–D3, mirrored in .claude/SCHEMA_DECISIONS.md). Analyzer 1.5.0; Neo4j contract stays 2.0.0.

What was wrong

span.bytes was produced (and documented in schema.ts) as UTF-16 char offsets — ts-morph's getStart()/getEnd() — while the canonical keystone and codeanalyzer-python mean UTF-8 bytes (byte_offsets, _span_code = source.encode()[lo:hi]). The projection's spanCode already sliced a Buffer by them, so :TSCallable.code ran short by exactly the multibyte surplus inside the span: python-sdk's sample src/index.main holds one em dash (3 bytes) → 2 bytes short → the closing \n} gone.

What changes

• span.bytes = UTF-8 byte offsets on every node at every level: declarations, call/config_access (L1), statements and @entry/@exit (L3), the module span ([0, byteLength]), yaml ConfigKey spans. Buffer.from(source).subarray(lo, hi) reproduces the text; source.slice no longer does on a non-ASCII file. Values change only for nodes after a non-ASCII character; ASCII files are byte-identical before and after (-j1/-j4 determinism re-checked).
• One utility, src/schema/offsets.ts: ASCII fast path is the identity; otherwise one cumulative table per text, cached per owning object (SourceFile / TSModule). Producers convert on the way out; the three consumers that need compiler positions (configUse.nodeAtSpan, the defuse linker's factory lookup, entrypoints/matching default-export resolution) convert on the way in. The dataflow IR keeps char offsets internally.
• Docs: schema.ts comments, the consumer skill's vocabulary, CLAUDE.md.

Tests

test/byte-offsets.test.ts: a fixture with an em dash, CJK and an emoji before and inside declarations — every callable, field, call site, config_access, L3 statement, @entry, the module span, the yaml key span, and the graph's code slice byte-exact; the config-use literal tier still resolves past multibyte text; offsetMapOf round-trips incl. surrogate pairs. schema-v2.test.ts's slice assertions now use the byte rule. bun test: 339 pass.

Downstream

python-sdk #343 (leg 2.5b): slice span.bytes as bytes on the TypeScript leg too — its python side already does — and the four xfail(strict=True) marks on test_typescript_bulk_parity_live.py flip when the pin moves to the release carrying this.