feat(annotation): configurable per-field markdown rendering

[henrycgbaker]

Summary

Adds AnnotationSettings.field_render_mode (deployment / workspace / task scopes, sparse-dict overlay) controlling use_markdown on each rg.TextField. Lets a deployment opt selected fields into Argilla's markdown renderer — which also handles inline raw HTML — so content from pipelines that emit HTML (e.g. publikationsbot) displays as HTML in the annotator UI rather than escaping.

Mirrors the constraint_severity pattern from #216 very closely:

	constraint_severity	field_render_mode
Shape	dict[str, Severity]	dict[str, FieldRenderMode]
Default factory	populated with all constraint IDs	populated with all known field names → "plain"
Merge validator	overlays user values onto defaults	overlays user values onto defaults
Key validator	rejects unknown constraint IDs	rejects unknown field names (scope-aware)
Resolver method	resolved_severity(ws, constraint_id)	resolved_render_mode(ws, task, field_name)
Scope	deployment + workspace	deployment + workspace + task

The one difference: field_render_mode adds a per-task scope (TaskSettings.field_render_mode) since "answer" belongs to different tasks (grounding vs. generation) and a deployment might legitimately want them rendered differently. constraint_severity skips per-task because severity is per-constraint, not per-task — that reason doesn't apply here.

Files changed

• core/schemas/annotation_task.py — adds FieldRenderMode type + TEXT_FIELDS registry
• core/settings/annotation_settings.py — adds field_render_mode to all 3 scope classes, plus resolved_render_mode(), _merge_field_render_mode_defaults, _validate_field_render_mode_keys
• core/annotation/argilla_task_definitions.py — local md() helper inside build_task_settings, wires use_markdown=md(task, name) into the 6 rg.TextField calls
• docs/decisions/0009-annotation-schema-configurability.md — addendum: rendering is operational, not schema (in the same layer as min_submitted, constraint_severity, locale); the hardcoded constraint on fields/questions/labels still holds
• Tests:
  • tests/unit/core/settings/test_annotation_settings.py — TestFieldRenderModeDefaults, TestResolvedRenderMode, TestWorkspaceFieldRenderModeValidation, TestTaskFieldRenderModeValidation
  • tests/unit/core/annotation/test_argilla_task_definitions.py — TestTextFieldsRegistryMatchesActualTextFields (drift check), TestUseMarkdownWireUp (propagation across 3 scopes)

Example YAML

# Deployment-wide: render HTML answers as markdown
field_render_mode:
  answer: markdown

# Or with per-workspace / per-task overrides:
workspaces:
  grounding:
    field_render_mode:
      context_set: markdown
    tasks:
      grounding:
        field_render_mode:
          answer: plain  # override the deployment default just here

Dependency chain

Branched on feat/annotation-logical-constraints/05-export-summary-by-id (#217 head). Depends on #213 → #214 → #215 → #216 → #217 landing first. The work itself is independent of those changes' content but uses:

• build_task_settings(settings) signature from refactor(annotation): DRY build_task_settings via assemble() helper #215
• task_to_workspace() helper + sparse-dict-overlay pattern from feat(annotation): configurable constraint severity at deployment + workspace scopes #216

Test plan

• pytest tests/unit/ — 754 pass, no regressions
• Manual: on a fresh test deployment, set field_render_mode: {answer: markdown} in YAML, run pragmata annotation setup, verify resulting Argilla datasets have use_markdown=True on grounding/generation answer fields (already demoed manually on a downstream demo branch).