feat: migrate parser.rs to PRS-0xx structured error codes — #277 (2/6) - #366

Summary

Step 2 of the 6-PR #277 rollout (structured runtime error codes). The
foundation PR (#357) built the machinery (MinigrafError, ErrorCategory,
ErrorCode, REGISTRY, bail_coded!/err_coded!) with zero call sites
migrated. This PR wires up the PRS (parser) category — the largest, at 79
codes — in src/query/datalog/parser.rs.

• Converts parser.rs's internal Result<_, String> signatures (118 error
  sites across ~30 functions) to anyhow::Result<_>, per the decision
  recorded in the foundation PR (parser.rs couldn't use
  bail_coded!/err_coded! as a drop-in without this — they produce
  anyhow::Error).
• Wires all 79 documented PRS-0xx codes to their call site via
  bail_coded!/err_coded!, matched by content against
  docs/ERROR_REFERENCE.md.
• src/db.rs's three parse_datalog_command() call sites
  (Minigraf::execute, WriteTransaction::execute, Minigraf::prepare)
  drop their .map_err(|e| anyhow::anyhow!("{}", e)) — that stringified and
  discarded any CodedError chain before MinigrafError::from could recover
  it at the API boundary.
• docs/ERROR_REFERENCE.md's PRS "Error text" lines now show the real
  {}-templated form (matching the INT-000 convention already
  established) instead of a fully-rendered example value — verified
  byte-for-byte against error.rs's REGISTRY by the existing
  registry_is_a_subset_of_error_reference_doc sync test.
• Two static message strings changed wording to match their documented
  "Error text" verbatim (a pre-existing drift): "Retract argument must be a vector" → "...must be a vector of facts" (PRS-044).

Not in scope / left uncoded: ~40 parser error sites have no documented
PRS code (defensive/unreachable internal errors, :max-derived-facts/
:max-results validation, a handful of (not-join)/(or) structural
messages) — these still fall back to INT-000 via MinigrafError::from,
unchanged from the foundation PR's behavior. Two documented codes are wired
up but structurally unreachable through the public API: PRS-028 ("window
expression cannot be empty") and PRS-049 ("unexpected end of fact
vector") — the has_over/len() >= 4 conditions guarding their call sites
already rule out the case they check for. Pre-existing parser logic, not
introduced by this change.

Test plan

• cargo test — 1097 passed, 0 failed (was 1081 on main; +16 new test
  functions in tests/error_codes_prs_test.rs, covering all 77
  reachable PRS codes end-to-end through db.execute() and
  db.prepare(), asserting both .code() and .category())
• Updated tests/error_codes_foundation_test.rs's placeholder assertion
  (execute_parse_error_returns_coded_error) from INT-000/Internal
  to the real PRS-005/Parser it now produces
• cargo clippy --all-features -- -D warnings — clean (matches
  .github/workflows/rust-clippy.yml's exact invocation)
• cargo fmt --check — clean
• Pre-push hook (fmt + clippy + test) passed

Does not close #277 — the umbrella issue for all 6 PRs stays open until
QRY, STG, WAL, and API land too.

Closes #358

Co-Authored-By: Claude Sonnet 5 noreply@anthropic.com

https://claude.ai/code/session_01P7e4uTBWtM8fMZiPZc4bCw