refactor: first-class tabular redaction + Strategy/DefaultStrategy merge + test cleanup

[martsokha]

Summary

Six commits across three threads, all atop the location/read/redact codec API from #148.

1. First-class tabular redaction (`211f95f`)

Spreadsheets get their own modality. CSV no longer pretends to be text; XLSX no longer pretends to be CSV.

• New `TabularHandler` capability trait + `BoxedTabularHandler` + `ContentHandle::Tabular` variant. Mirrors the text/image/audio shape.
• CSV implements `TabularHandler` only — cells addressed by `(row_index, column_index)`. Row 0 = headers if present; intra-cell `start_offset`/`end_offset` optional (omitted = redact whole cell). The old byte-offset text path with its ~250 lines of `locate_cells` / `find_field_in_line` machinery is gone.
• XLSX gets a tabular-only stub (real implementation pending).
• Engine `RedactionApplicator` gains `build_tabular_redactions` parallel to text/image/audio. `Document::value_at(Location::Tabular)` now actually reads the cell.

Folder layout matches modality: `handler/text/` no longer holds spreadsheet handlers — they live in `handler/tabular/`.

Redaction code-review fixes folded in:

• `entity_map` + `mapping_index` built once at top of `apply()` (was rebuilt per phase with O(n²) scans).
• `chars().count()` in mask repeat (multi-byte grapheme bug).
• Per-strategy audit placeholders: `[IMAGE_BLUR:5]` / `[IMAGE_BLOCK:#ff0000]` / `[AUDIO_SILENCE]` instead of a single `[IMAGE_REDACTED]` for everything.
• Explicit per-variant strategy → output dispatch; `#[non_exhaustive]` fallthroughs warn-log instead of silently producing a generic placeholder.

2. Strategy / DefaultStrategy merge (`b0a59d3`)

`Strategy` and `DefaultStrategy` were structurally the same idea — "what to do per modality" — represented as two different types: an enum that picks one variant, and a struct that allows any combination. `DefaultStrategy::for_location` was the bridge translating one to the other.

Collapsed into a single `Strategy` struct with `Option<*Strategy>` per modality. Each per-modality enum gains `Default`:

• `TextStrategy::default()` = `Replace { placeholder: "" }` (engine fills `[ENTITY_KIND]`)
• `ImageStrategy::default()` = `Block { color: black }`
• `AudioStrategy::default()` = `Silence`

`text_or_default`/`image_or_default`/`audio_or_default` accessors resolve to a concrete strategy. The applicator already filters by `entity.location` variant, so it calls the matching accessor directly — no enum dispatch needed.

`log_modality_mismatch` is gone. Every modality always resolves to a concrete strategy now — the concept of "rule wants text but entity is image" no longer exists. A rule that only specified an image strategy on a text entity now resolves to `TextStrategy::default()` and produces `[ENTITY_KIND]`.

`Strategy::is_reversible_for(&Location)` replaces the old `is_reversible()` — reversibility is inherently per-modality.

3. Test cleanup (4 commits)

134 tests removed across the workspace. Each fell into one of:

• Tests derives: serde round-trip on simple structs, `Display` from `derive_more` or `strum`, `Default` derives, `Clone`/`Copy` checks.
• Tests dependencies: `oxilangtag` wrapper, `strum` derives, `uuid` v7 generator, `infer` crate, the `csv` crate's quoting.
• Trivial getters/constructor read-backs: `Location::as_text` variants, `builder_required_fields`, `ContentMetadata` accessors, http_error builder-set-fields tests, PDF `accessors`/`empty_handler`, etc.
• Duplicates of nearby tests: `pattern_expression_roundtrip_glob` (duplicates `_regex`), `no_overlap_different_col` (duplicates `_row`), `load_detects_four_space_indent` (duplicates `two_space`).
• Tests stdlib behavior: `Vec::is_empty` after `new()`, `Vec::find` returning `None`, `HashMap::contains_key` after one `insert`.
• Registry-has-entries smoke checks the user explicitly flagged as worthless (`builtins_load`, `all_error_kinds_have_responses`, `resolver_builtins`).

Plus `DataReference` deleted (`da3cdb5`) — unused type with no callers anywhere in the workspace.

Per-crate test count deltas:

• nvisy-codec: 93 → 63
• nvisy-core: 25 → 11
• nvisy-ontology: 120 → 65
• nvisy-pattern: 66 → 57
• nvisy-engine: 85 → 72
• nvisy-server: 22 → 12

Commit walkthrough

• `211f95f` — first-class tabular redaction (codec + engine, ~1244+/-711 across 17 files)
• `981faf6` — codec test cleanup (-30 tests / -297 lines)
• `f8ac8e1` — core/ontology test cleanup (-69 tests / -656 lines)
• `da3cdb5` — drop unused DataReference (-66 lines)
• `3d4e685` — pattern/provider/engine/server test cleanup (-35 tests / -616 lines)
• `b0a59d3` — Strategy / DefaultStrategy merge (~297/-297 across 34 files)

Test plan

• `cargo check --workspace --all-features` clean
• `cargo clippy --workspace --all-features -- -D warnings` clean
• `cargo test --workspace --all-features` — all green
• `cargo doc --workspace --no-deps --all-features` — no new warnings

🤖 Generated with Claude Code