cardigann: parse templates once at load (parse-don't-validate lifecycle)

[forkwright]

Finding

The Cardigann engine parses every template twice: once at definition load for validation (template::validate / template::validate_row_scoped in crates/eksetasis/src/client/cardigann/template.rs), and again at render time (render_specs / the extractor path). The load-time functions return Result<(), _> and discard the parsed AST.

Evidence

• crates/eksetasis/src/client/cardigann/template.rs:108 (validate) and :116 (validate_row_scoped) both call parse(template)? then discard the nodes.
• Call sites in definition.rs:365,371,381,383 use them purely as validators; the definition struct stores raw template strings.
• RUST/validate-returns-unit fires on validate_row_scoped (the validate_ prefix); the sibling validate escapes the pattern but has the same shallow shape.

Why this matters

Parse-don't-validate (RUST.md#validation-constructors): a parsed template is the constrained type. Storing parsed ASTs in the definition would make load-time validation produce the value the render path consumes — one parse, no re-parse per search, and invalid states unrepresentable past load. This is the same class as aitesis#301.

Desired correction

Refactor the definition lifecycle: CardigannDefinition stores parsed templates (or a ParsedTemplate newtype over the node AST) produced once at load; the render path consumes them. validate/validate_row_scoped fold into the parser.

Done when: templates parse exactly once per definition load; no validate_* function returns Result<(), _>; the .kanon-lint-ignore entry added under #513 is removed.

[forkwright]

State checkpoint (weekly usage window): implementation in progress on branch refactor/cardigann-parse-once (worktree preserved at ~/dispatch-wt/harmonia-696). A clearly-marked UNVERIFIED WIP commit captures the mid-refactor state of template.rs + definition/yaml.rs. Resume: read the diff, run cargo test -p eksetasis, complete the lifecycle, then remove the .kanon-lint-ignore validate-returns-unit entry per this issue's done-when.

[forkwright]

There is unlanded work for this on a branch, and nothing points at it. Recording it here so it is discoverable — the seat that wrote it has ended.

forkwright/harmonia branch refactor/cardigann-parse-once, last commit wip(eksetasis): parse-once cardigann lifecycle at 2026-08-12. No PR was ever opened.

Against origin/main it is 2 files, +217/-190:

crates/eksetasis/src/client/cardigann/definition/yaml.rs
crates/eksetasis/src/client/cardigann/template.rs

Both differ from main, and the branch adds no new files — so no path-existence check will show it as unlanded. The commit is marked wip, so it is not claimed finished.

A branch with no PR, no issue reference, and modification-only content is invisible to every sweep that adjudicates by "does this branch add something main lacks." I checked this repo today while driving it toward a single branch; this is why it still has two.

Source: first-hand — git diff --shortstat origin/main...origin/refactor/cardigann-parse-once and a per-file comparison against origin/main, run today.