Skip to content

fix!: v0.4.0 bug batch — fences, whitespace, type safety, @extends FM (#149–#154)#161

Open
dean0x wants to merge 19 commits into
mainfrom
fix/v0-4-0-149-154
Open

fix!: v0.4.0 bug batch — fences, whitespace, type safety, @extends FM (#149–#154)#161
dean0x wants to merge 19 commits into
mainfrom
fix/v0-4-0-149-154

Conversation

@dean0x

@dean0x dean0x commented Jul 8, 2026 •
edited
Loading

Copy link
Copy Markdown
Owner

Summary

Changes

Core behavior (BREAKING)

Gap-closure (post-review)

Breaking Changes

Reviewer Focus Areas

  • crates/mds-core/src/resolver/frontmatter.rs — deep_merge_yaml depth-gating logic (depth == 0 guard) and unit tests
  • crates/mds-core/src/lib.rs — clean_output new implementation (trim only)
  • crates/mds-core/src/lexer.rs — indented/tilde/blockquoted fence recognition
  • crates/mds-python/tests/test_parity.py — golden updated for interior-verbatim contract (blank line after FM fence now preserved)

Closes #149
Closes #150
Closes #151
Closes #152
Closes #153
Closes #154

dean0x and others added 19 commits July 8, 2026 19:03
@dean0x @claude
fix: interpolation error hint recommends \{ (#153)
a449a50 
The hint in the invalid-interpolation error message was rendering `\{{`
(double brace) instead of `\{` (single brace) because the Rust format
string used `\\{{{{` instead of `\\{{`.

Fix the format string and add a regression test that pins the rendered
hint text. Also include the v0.4.0 design artifact for issues #149-#154.

Co-Authored-By: Claude <noreply@anthropic.com>
@dean0x
fix!: interior-verbatim whitespace contract (#150, #151)
70359f7 
Delete-directive-lines model: a directive line consumes exactly its own
line + trailing newline; every other markdown-mode byte passes through
verbatim. Replaces the old leading-trim + blank-run-cap approach.

Core changes:
- clean_output: rewritten to trailing-edge normalisation only (strip \r,
  trim trailing whitespace, ensure one final \n); interior content passes
  through byte-exact. Rename/update unit tests.
- parser: remove strip_leading_newline / strip_trailing_newline calls from
  @block and @define parse sites; delete both dead functions and module-doc
  references. @define call site now trims via evaluator instead.
- evaluator: invoke_function now calls .trim() on the body evaluation result
  at the call site (replaces parser-level edge strip for @define).
- formatter: delete R3 (blank-line run capping) — now mismatches the
  new clean_output and would break compile-equivalence. Also deletes
  protected_spans() + all threading (code fences no longer need a separate
  protection class; without R3 they get the same R1-only treatment as
  ordinary body). Module doc updated to describe v2 interior-verbatim
  ruleset. doctest expected value updated.
- All affected tests updated to reflect preserved blank-line runs and
  verbatim block body trailing \n.

Adds repro integration tests: #150 (blank-line run in compiled output)
and #151 (block body edge newline visible at compile boundary).

BREAKING: blank-line runs and block-body edge newlines now pass through
to compiled output. Templates that relied on run-capping or edge stripping
will produce different markdown.
@dean0x @claude
fix: recognize indented, tilde, and blockquoted code fences (#149)
a96436c 
- Widen fence recognition to accept [ \t>]* prefix before the fence
  characters, matching CommonMark indented/blockquoted fences
- Add tilde (~~~) fence support alongside backtick (```) fences
- Track open fence as Option<(char, usize, usize)> (char, count, opener
  byte offset) instead of a bare count; closing fence must match the
  opening character and have count >= opener count
- Use opener byte offset in the unclosed-fence error so the diagnostic
  spans point to the opening line rather than end-of-file
- Add 6 lexer unit tests covering tilde, indented, blockquoted, char
  mismatch, count mismatch, and unclosed-fence span

Co-Authored-By: Claude <noreply@anthropic.com>
@dean0x @claude
fix!: cross-type comparison errors, structural equality, --set-string (…
b57e2fd 
…#152)

BREAKING: @if conditions that compare values of different types (number vs
string, boolean vs null, etc.) are now a TypeMismatch runtime error instead
of silently returning false (==) or true (!=). Templates relying on
cross-type comparisons must add explicit str() or num() conversions.

Changes:
- Rename values_equal_runtime → values_equal_same_type; return Option<bool>
  (None = cross-type, caller emits TypeMismatch error)
- Add MdsError::TypeMismatch with code mds::type_mismatch and a help hint
- Update evaluate_condition Eq/NotEq branches to propagate TypeMismatch
- Add RuntimeVarArgs struct to bundle --vars/--set/--set-string together
- Add --set-string KEY=VALUE flag to build, check, and watch subcommands:
  forces value to remain a string with no type coercion
- Update language.rs tests: cross-type comparisons now expect errors
- Add virtual_fs.rs issue repros for TypeMismatch (#152) and --set-string

Co-Authored-By: Claude <noreply@anthropic.com>
@dean0x
fix!: @extends emits deep-merged frontmatter (#154)
d099b25 
Previously the compiled output of a child template only contained the
child's raw frontmatter, silently discarding any keys inherited from
the base.  The fix serializes the deep-merged mapping (base < child,
reserved keys `imports`/`type`/`extends` excluded) produced by
`build_merged_extends_scope` and uses it as the output frontmatter for
both the text-cached (`process_module_extends`) and intrinsic
(`process_module_intrinsic`) compile paths.

Breaking: `@extends` children now emit base frontmatter keys in their
compiled output.  Templates that relied on base keys being invisible in
the output must strip them explicitly.

Implementation:
- `build_merged_extends_scope` returns `(Scope, Mapping)` — the mapping
  is no longer dropped after scope construction.
- `serialize_merged_frontmatter(mapping)` → `Option<String>` serializes
  the mapping with `serde_yaml_ng`; returns `None` when empty.
- `ExtendsComponents` gains a `merged_frontmatter: Option<String>` field
  populated by `resolve_extends_components`.
- Both terminal paths use `merged_frontmatter` instead of the child's
  raw frontmatter string.
- `process_module_extends` no longer receives `raw_frontmatter` as a
  parameter (it is now derived from the merged mapping).
- `f4_child_emits_only_own_raw_frontmatter` rewritten as
  `f4_extends_emits_deep_merged_frontmatter` to match the new contract.
- 4 new `virtual_fs` integration tests (issue_154_*).

Closes #154
@dean0x
docs: v0.4.0 language/CLI docs, changelog, examples
63c2958 
spec.md (bumped to v0.4):
- Code fence passthrough now mentions tilde, indented, and blockquoted
  variants (#149)
- Equality strict rule updated: cross-type comparison is a runtime
  error (mds::type_mismatch), not false/true (#152)
- Frontmatter merging rule updated: @extends emits the deep-merged
  mapping (base < child), not just the child's raw FM (#154)
- `--set-string` added to CLI options table (#152)
- `mds fmt` added to commands table and §7.4 section (#60)
- CLI sections renumbered (7.4–7.8) after inserting the fmt entry
- §12 Status updated with v0.4.0, v0.3.0 entries; spec version bumped

CHANGELOG.md:
- [0.4.0] Unreleased section with BREAKING notices for #152 and #154,
  Added (--set-string, mds fmt, Python bindings), Changed (#146),
  Fixed (#149, #153, #133/#146)

Closes #149
Closes #150
Closes #151
Closes #152
Closes #153
Closes #154
@dean0x
fix!: deep_merge_yaml gates reserved-key filter on depth == 0
8226313 
Reserved YAML keys (imports, type, extends) must only be stripped at
the top level of the merged frontmatter mapping. Previously the guard
was unconditional, silently dropping nested keys like config.type or
config.imports that are user data, not directives.

Change: both Phase 1 and Phase 2 of deep_merge_yaml now check
`depth == 0` before testing RESERVED membership. Recursive calls
(depth > 0) pass nested keys through verbatim.

Tests added:
- deep_merge_yaml_strips_top_level_reserved_keys (unit)
- deep_merge_yaml_nested_reserved_keys_preserved (unit, GAP 1)
- deep_merge_yaml_both_empty (unit)
- f6_nested_reserved_key_config_type_survives_merge (integration)

Closes #154 (nested key correctness sub-case)
@dean0x
test: adversarial frontmatter YAML injection CLI mirror tests
ac3336b 
Add two adversarial tests in crates/mds-cli/tests/frontmatter.rs
that verify YAML values cannot escape the frontmatter fence block
and inject extra top-level keys into compiled output.

Tests:
- adversarial_value_containing_fence_sequence_via_extends: a base
  frontmatter value containing literal "\n---\n" goes through the
  @extends re-serialization path; asserts the injected key does not
  appear as a top-level frontmatter key (uses extract_frontmatter
  helper to check only the FM section, not body text).
- adversarial_block_scalar_with_type_mds_line: a block-scalar value
  whose lines include "type: mds" must not be confused with the
  top-level type: mds directive; only the unindented top-level key
  is stripped.

Also adds extract_frontmatter() helper to check only the YAML between
the first ---/--- pair, avoiding false positives from --- in body text.
@dean0x
docs: README --set-string flag, mds fmt interior-verbatim contract, s…
da1a943 
…pec v0.4

- Add --set-string KEY=VALUE to Build/Watch flags table (it was missing
  alongside --set); both flags ship in v0.4.0
- Update mds fmt "What it normalizes" section: remove the stale
  "two or more consecutive blank lines collapse" bullet — R3 was removed
  in v0.4 (#150/#151); the formatter now passes blank lines verbatim
  under the interior-verbatim contract
- Bump inline spec version reference from v0.2.0 to v0.4.0
@dean0x
docs: add fence variants edge-case example (#149)
26eead1 
Add examples/edge-cases/26_indented_and_tilde_fences.mds demonstrating
the four code-fence forms now recognized by the compiler:
1. Standard backtick fence (baseline)
2. Tilde fence (~~~rust ... ~~~)
3. Indented fence inside a list item (4-space prefix)
4. Blockquoted fence (> ``` ... ```)

All four suppress interpolation inside the fence. Interpolation resumes
normally after each closing fence. Literal {x}/{name} preserved inside,
{language} interpolated outside.

Closes #149 (example coverage)
@dean0x
test(python): update parity golden for interior-verbatim contract (#150)
7fe1a6b 
The frontmatter golden had the pre-#150 expected output: a blank line
between the closing --- fence and the body text was silently stripped
by the old clean_output leading-blank-line behavior.

Under the interior-verbatim contract (#150/#151), clean_output no longer
strips leading blank lines — it only does trim_end() + forced final \n.
The golden now reflects what both the CLI and the Python binding produce:
`---\n\nHello Alice!` (blank line preserved verbatim).
@dean0x
refactor: post-implementation simplification pass
ccd1a70 
- lexer: remove unused _prefix_len parameter from scan_code_fence;
  replace if/else-if-let/else with match to eliminate unreachable
  else branch (Option is exhaustive); drop stale doc line referencing
  the removed parameter
- evaluator: rename section header comment from values_equal_runtime
  to values_equal_same_type (tombstone name residue from rename)
- resolver: fix blended doc comment — serialize_merged_frontmatter was
  inserted before parse_frontmatter_mapping leaving the former's doc
  concatenated onto the latter's old header; split cleanly and restore
  parse_frontmatter_mapping's own doc comment
@dean0x
chore: sync package-lock workspace versions to 0.3.0
324b6e5
@dean0x
fix: propagate merged-frontmatter YAML serialization errors
93e0b38 
serialize_merged_frontmatter previously swallowed serde_yaml_ng::to_string
failures via an Err(_) => None arm, which would silently drop the deep-merged
@extends frontmatter fence from compiled output on failure. Return
Result<Option<String>, MdsError> and map the error to MdsError::yaml_error so
the failure surfaces as a compile error at the boundary (no swallowed errors).
Thread the ? to the single caller in resolve_extends_components.

Also drop two stale R1/R3 comments in formatter.rs left over from the R3
removal (#150/#151) — R3 no longer exists; only R1 remains for body content.
@dean0x @claude
fix!: reject cross-flag duplicate keys across --set/--set-string (#152)
6fcf315 
Any key present in both --set and --set-string is now a hard CLI error
with message: "variable '{key}' is set by both --set and --set-string;
use only one". Same-flag last-wins is preserved unchanged.

Also add structural equality for same-type container comparisons in
values_equal_same_type: Array==Array and Object==Object now return
Some(bool) rather than falling through to _ => None (TypeMismatch).
Value derives PartialEq so the comparison is free. Cross-type remains
None (error). A short comment guards future grammar extensions that may
surface container literals on the RHS.

Co-Authored-By: Claude <noreply@anthropic.com>
@dean0x @claude
test: alignment-mandated coverage (round-trip, injection, repros)
6a8b517 
Adds:
- Core-level adversarial YAML-injection tests for @extends merged-FM
  serialization (fence sequences, YAML end markers, block-scalar type:mds)
  in resolver_tests.rs — core twin of the CLI mirror in frontmatter.rs
- Round-trip acceptance test: realistic template (frontmatter, post-FM
  blank, lists, indented fence in list item, interior blank runs,
  trailing-space line) compiled through a pass-through; asserts
  byte-identical output modulo the two carve-outs (trailing-edge
  normalization, \r strip) per the interior-verbatim contract (#150/#151)
- Integration repro for #149: indented fence inside list item with
  literal {braces} → compiles, braces literal, fence lines verbatim
- Integration repro for #153: invalid interpolation → error hint
  contains \{ and NOT \{{
- f7 extension: emitted raw_frontmatter ignores --set runtime var whose
  key collides with a frontmatter key (body uses runtime value, FM does not)

Co-Authored-By: Claude <noreply@anthropic.com>
@dean0x @claude
docs: interior-verbatim whitespace contract in spec + changelog (#150, …
a3b7b2e 
…#151)

spec.md:
- §4.2: Add fence out-of-scope note (4-space indent not recognized,
  blockquoted fence requires explicit closer)
- §4.3: Add comparison semantics table (types × ==/!= outcome incl.
  cross-type error, truthiness-vs-equality split, --set-string footgun)
- §4.11: Fix worked example output to match actual compiler (blank lines
  from base skeleton are now present in output)
- §4.11: Rewrite whitespace contract to interior-verbatim with
  trailing-edge normalization (removes stale "block body edges are
  stripped" claim; distinguishes @message/@define edge-trim from block
  body interior-verbatim)
- §4.11: Add named asymmetry callout (@extends emits canonically
  re-serialized merged frontmatter vs standalone raw-verbatim)
- §7.4: Remove false "collapses blank lines to one" claim from mds fmt
  description; re-pitch as line-ending/trailing-whitespace normalizer
- §12: Add #150/#151 interior-verbatim contract to v0.4.0 status entry

CHANGELOG.md:
- Remove stale ## [0.4.0] — unreleased heading; all content moves
  directly under ## [Unreleased] for bump-version.mjs compatibility
- Add BREAKING bullet for #150/#151 interior-verbatim whitespace
  contract under Changed, with migration note
- Fix --set-string Added bullet: was "mds fmt" (wrong), now "mds watch"
- Fix mds fmt (#60) Added bullet: remove blank-line collapsing claim

Co-Authored-By: Claude <noreply@anthropic.com>
@dean0x @claude
fix: type_mismatch help per contract + stale fmt help text (#152, #150)
c115bc2 
- error.rs: Replace TypeMismatch help text with contract-specified wording
  interpolating lhs_type/rhs_type and referencing --set-string KEY=VALUE
  and `@if x:` for truthiness; remove doc-comment references to str()/num()
  conversion functions that do not exist in the MDS language.
- virtual_fs.rs: Add issue_152_type_mismatch_help_mentions_set_string test
  that pins the --set-string mention via miette::Diagnostic::help so the
  contract stays locked and cannot silently regress.
- main.rs: Rewrite Fmt subcommand doc/help to match reality after R3
  removal (#150/#151) — drop the stale "collapses runs of 2+ consecutive
  blank lines" claim; describe actual behaviour (LF normalization, trailing
  whitespace on directive lines, final newline, safety gate).

Co-Authored-By: Claude <noreply@anthropic.com>
@dean0x @claude
ci: raise WASM soft-bloat guard 650K→700K (#149)
09a6096 
v0.4.0 code additions (cross-type comparison errors, extended help
strings, @extends frontmatter merging) plus ongoing `stable` toolchain
drift pushed the optimised binary to 650,745 bytes in CI (650,675 bytes
locally with slightly different wasm-opt). Both exceed the old 650K cap.

Raise the threshold to 700K and document the full budget history in the
comment. All 42 wasm-pack tests pass; guard is advisory soft-bloat, not
a functional gate.

Follow-up: pin the wasm toolchain to make binary size deterministic and
re-tighten the guard.

Co-Authored-By: Claude <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment