fix(parser)!: preserve contextual literal substitutions

[e54-bot]

Summary

Catalog coercion facts now decide only which literal substitutions a parameter accepts; the parser keeps the authored literal. Parse → emit no longer rewrites False/True → 0/1, 0/Vector(0, 0, 0) → Null, or Empty Array → "".

• Removes the parse-time normalization (normalize_contextual_argument, normalize_modify_value, normalize_modify_call_value) from value, assignment, and modify parsing. Validation already accepted the authored forms via contextual_value_matches, so acceptance is unchanged.
• ADR-0015 records the decision (option 2 of Decide whether parsing may discard OverPy optimizeForSize literal substitutions #298) and supersedes ADR-0011 decision 2. Emission-side size optimization (option 3) is rejected: optimizeForSize stays opy-rs policy, which can now write the substituted form into Program itself.
• docs/element-count.md records the evidence gap: no client capture establishes whether a numeric literal and False/True/Null cost the same. The model is unchanged; the analysis now sees the authored literal and can apply whatever rule a capture establishes.

Behavior change

Program and round-trip equivalence now distinguish 0 from False, 1 from True, and 0/zero vector from Null at coerced positions (marked breaking). Consumers that compared against normalized output (e.g. opy-rs reference comparisons) adopt this in their own PRs.

Client evidence

None available: no capture of the element cost of 0 versus False in a numeric argument exists. The ADR-0005 capture is follow-up work and does not change this decision.

Verification

• cargo fmt --all --check, cargo clippy --workspace --all-targets -- -D warnings, cargo test --workspace --all-targets (all pass), workshop-catalog-gen -- check, git diff --check.
• New contextual_aliases_survive_parse_and_emit_in_each_locale: the Decide whether parsing may discard OverPy optimizeForSize literal substitutions #298 table spellings survive en-US parse → emit verbatim, and a zh-CN emit → parse → emit is stable and equivalent to the authored program.
• Existing contextual-semantics tests now assert the authored value is preserved at each coerced position (actions, vector components, operators, conditional, array index, indexed/modify contexts).
• Not run: the motivating OWBastion/Bastion Wright build comparison, which needs the opy-rs adoption.

Fixes #298

[Teakowa]

LGTM