Beislið v0.4.10

Beislið v0.4.10

This release covers all changes since the last published GitHub release, v0.4.0.

Highlights

• The workflow.md parser contract is now canonical: a golden conformance corpus (17 cases), a frozen error-code catalog, and YAML-lite scalar semantics, all running in CI (BEI-131).
  This is the port target for olin - implementations port against tests/conformance/cases/, not against the Python code.
• P0 integrity fixes: fail-closed action-policy payload validation, run-ledger redaction aligned with policy secret patterns, and interrupt-safe staged project installs (BEI-128, BEI-129, BEI-130).
• Execution-envelope contract extracted into declarative JSON schemas, orphaned validators wired into CI, and a nightly agent-smoke suite (BEI-133, BEI-134, BEI-135).

Parser behavior changes (BEI-131)

The deterministic decisions from the 2026-07-03 HITL session (docs/plans/2026-07-03-bei-131-canonical-parser-decisions.md) are now implemented and pinned by the corpus.
Each of these is a behavior change from v0.4.9 and earlier:

• Trailing comments in unquoted scalars are stripped (YAML semantics): a # preceded by whitespace ends the value, so parallel_safe: true # optional now parses as boolean true.
  Quoted values keep # literally.
  Previously the comment stayed inside the value.
• Inline flow maps are rejected: - { name: lint, command: 'pnpm lint' } is now a flow_map_unsupported error instead of a silent literal string.
  Block style is the one canonical mapping syntax.
• Parse diagnostics report absolute workflow.md line numbers, not indexes into a comment-stripped array.
• Unknown beislid:* fence keys now warn (unknown_fence_key, exit 0) against the full documented fence-key registry (~25 keys), instead of being silently ignored.
  Newer fence keys degrade gracefully against older normalizers.
• Registered-but-unconsumed fence blocks are now syntax-checked: a malformed scopes or babysit block fails the file with a parse error, where the old normalizer never looked at it.
  Well-formed blocks owned by other tools are still skipped silently.
• Tier enums are validated at normalize time: tiers keys must be light, standard, heavy, or frontier, and tier_mode must be prefer or require; violations are invalid_value errors.
• model and models on one route are mutually exclusive: both present is now an invalid_value error instead of model being silently dropped.
• Decimal floats parse as numbers: -?\d+\.\d+ (e.g. 1.5) is a JSON number.
  Exponents, 1., .5, and 1.5.2 stay strings.
• Nested inline lists are rejected: a [ inside an inline list is a nested_inline_list error pointing at block style.
• Strict quote semantics: double-quoted scalars process exactly \n, \t, \\, \"; any other escape is an unknown_escape error.
  Single-quoted scalars are literal except '' -> '.
  Unterminated or mismatched quotes are unterminated_quote errors instead of silently passing through as literal strings.
• Unterminated fence blocks are malformed_block errors: a beislid:* fence without a closing fence line was previously ignored entirely; it now fails the file.
• Unterminated inline lists are malformed_block errors (a missing ] with no trailing comment).
• gate_sets and lifecycle_actions pass through unknown keys in the normalized envelope instead of projecting down to their known fields, matching the other sections.
  Note: sets: null is no longer coerced to {}.
• Duplicate beislid:* fences keep the existing behavior (warn + first occurrence wins), now pinned by the corpus.

Parser infrastructure (BEI-131)

• Golden conformance corpus: tests/conformance/cases/ (17 input/expected-envelope pairs) with the scripts/run_conformance.py runner, wired into CI as the normalizer-conformance job.
• Frozen diagnostic catalog: docs/parser-conformance.md documents all 14 error/warning codes with severity and path conventions.
  These codes are stable from this release onward.
• Format doc amendments: the scopes examples use block style, and the YAML-lite scalar rules (comments, floats, quotes, inline lists) are documented in .beislid/workflow-md-format.md, so doc and implementation agree.

Fixes and hardening

• Fail-closed action-policy payload schema validation (BEI-128).
• Run-ledger redaction aligned with policy secret patterns (BEI-129).
• Interrupt-safe staged project copy installs (BEI-130).
• Orphaned validators and tests wired into CI (BEI-133).
• Skill portability gaps fixed for non-Claude hosts.

Workflow and orchestration

• Execution-envelope contract extracted into declarative JSON schemas (BEI-134).
• Nightly agent-smoke suite with fixture library and blueprint/implement/review/rinse/setup scenarios (BEI-135).
• command_proofs added as a distinct per-slice envelope contract field (RON-146).
• Babysit now requires a rondo.final_report/v0 JSON final response.
• Workflow-signal heartbeat hook, sweep command, and stale reporting.
• scripts/validate.sh local CI mirror plus docs/testing.md.