refactor(P3a): move per-category behavior flags into a ModalityRegistry

[LukasWodka]

Summary

Structural refactor — phase P3a (epic backend#796; design spike here). First slice of the ModalityRegistry — the single source of truth for per-category behavior.

This slice introduces the modalities/ package and migrates the three hand-maintained frozensets in ingestors/base.py onto it. Smallest, lowest-risk slice — it establishes the pattern; P3b (validators) and P3c (transfer plan) follow.

Changes

• modalities/spec.py — ModalitySpec dataclass. P3a carries the 3 behavior flags (is_file_bearing, is_tabular_family, is_self_supervised); it grows in P3b–P3d with the validator factory, the sidecar transfer plan, and the conventions defaults.
• modalities/registry.py — REGISTRY (one spec per category) + spec_for() + the three category sets derived from the specs, so they can't drift from the source.
• ingestors/base.py — deletes the 3 frozenset literals (~50 lines) and imports the derived sets under their previous names. Every category in <set> check — and its None/unknown → False semantics — is unchanged.

Why it's behaviour-preserving

The derived sets equal the old frozensets exactly: 7 file-bearing / 4 tabular / 1 self-supervised. base.py's logic didn't change — only where the sets come from.

• Full suite: 1033 passed, 1 xfailed, coverage 97.3%, modalities/ 100%.
• tests/test_category_congruence.py and tests/test_ingestor_base.py (which import _FILE_BEARING_CATEGORIES from base) pass unchanged — the aliased import keeps the name resolvable.
• The test(e2e): characterization harness pinning clean-ingest behavior per modality #247 characterization goldens are unaffected (no data-path change); CI e2e re-confirms on real MySQL.

The new invariant

tests/test_modality_registry.py pins registry ↔ TaskCategory ↔ schema-enum equality and that the derived sets match the spec flags. This is the structural upgrade over #240's congruence test: once P3 completes, the per-category data has one source the call sites are derived from — making the instance_segmentation half-wired-zombie class (#240 / #99) unrepresentable rather than test-caught.

Scope

P3a is flags only — map_validators, map_file_transfer, and the conventions groupings are untouched (migrated in P3b/P3c/P3d). No conflict with other work.

🤖 Generated with Claude Code

[LukasWodka]

👋 Heads-up — Code review queue is at 14 / 8

Above the WIP limit. The team convention is to review existing PRs before opening new work.

Open PRs currently in Code review (oldest first):

• averaging-service#106 — test(integration): pin XGBoost/LightGBM/CatBoost end-to-end averaging round-trips · author: @LukasWodka · no reviewer assigned
• averaging-service#111 — fix(averaging): correctness round-2 sweep — N5/N7/N8/N9/N10 + load_model except (Sync develop → master for v0.3.0 release #109) · author: @LukasWodka · no reviewer assigned
• client#254 — Sync develop → main for v1.7.1 chart release (egress-enforcement preflight helm test, inert) · author: @saadqbal · no reviewer assigned
• client#256 — fix(egress-proxy): probe reports DNS failure distinctly, not as a TCP connect [image_validator falsely rejects images when target_size matches exactly #104] · author: @saadqbal · no reviewer assigned
• client-runtime#105 — chore(ci): drop unused mysqlclient dep; report test coverage (Phase 0) · author: @LukasWodka · no reviewer assigned
• client-runtime#106 — fix(schema): sync vendored ingest.v1.json — restore token_classification · author: @LukasWodka · no reviewer assigned
• data-ingestors#252 — fix(ingest): catch single-label classification at preflight + surface backend reason (issue Single-label classification dataset surfaces as misleading 'Backend failed to prepare' instead of a clear preflight error #251) · author: @divyasinghds · no reviewer assigned
• design-system#47 — feat(table): Table compound organism (Figma 959:6061 + 986:9822) · author: @waqaskhanroghani · no reviewer assigned
• frontend-app#523 — test(cypress): session + protected-route + error-state + endurance E2E coverage (and 2 bug fixes) · author: @waqaskhanroghani · no reviewer assigned
• tracebloc-py-package#218 — fix(upload): output-shape validation + hard-fail class-count check for text_classification · author: @shujaatTracebloc · reviewer: @aptracebloc

Pull from review before opening new work. (This is a nudge from the kanban WIP check, not a block.)