Feat/explainer story i18n

[Creylay]

Summary

Adds a story() narrative hook to every explainer (13/13), so global and local explanations come with a deterministic, human-readable summary in addition to their plots. Stories are generated in all 5 supported languages (en, es, pt, de, zh) at once and returned alongside the plot response, so the frontend can switch languages instantly (via the existing i18n selector) without refetching. Also removes redundant hardcoded English text that a subset of explainers previously baked into their plots, now superseded by the multilingual story() output.

---

Type of Change

• Backend change
• Backend change
• Frontend change
• CI / Workflow change
• Build / Packaging change
• Bug fix
• Documentation

---

Changes (by file)

Core story infrastructure

• DashAI/back/explainability/story.py: new format_story()/concat_stories() helpers that build a MultilingualString from per-language templates, reused by every explainer instead of duplicating 5-language boilerplate.
• DashAI/back/explainability/global_explainer.py, local_explainer.py: added an optional story(explanation, explainer_output) hook to BaseGlobalExplainer/BaseLocalExplainer (default None), mirroring the existing plot() contract.

DashAI/back/explainability/explainers/ (Explainers)

• permutation_feature_importance.py, regression_permutation_feature_importance.py: new story() ranking features by importance; only claims the model "relies on" features with measurable (>0) importance, calling out separately when the rest showed none instead of lumping them together.
• partial_dependence.py, regression_partial_dependence.py: new story() classifying each curve as increasing/decreasing/flat/non-monotonic from its actual values; added a dedicated "flat" case for curves with no real change (previously misreported as "increases").
• kernel_shap.py: new story(); removed the old English-only Plotly annotation baked into the chart (predicted class/probability), now covered — with more detail — by story().
• contrastive_shap.py, dice_counterfactual.py, nearest_counterfactual.py, regression_kernel_shap.py, token_ablation.py, grad_cam.py, occlusion_saliency.py, lime_text.py: migrated each explainer's existing hardcoded English TextArtifact summary into a multilingual story() with equivalent wording, and removed the now-redundant TextArtifact from plot().

** DashAI/back/api/api_v1/endpoints/explainers.py (API)**

• GET /global/plot/{id} and GET /local/plot/{id} now also load explanation_path (previously only plot_path/plots_path) and attach a "story" key ({"en": ..., "es": ..., ...} or null) to each artifact/group in the response.
• Added _resolve_story_explainer() to reconstruct the explainer from its registered class + stored parameters (no trained model needed — story() never touches it) and _attach_stories()/_attach_one_story() to compute and attach stories per artifact, never raising: any failure (unbuildable explainer, mismatched shape) just leaves "story": null.
• Added _as_group_target()/_as_artifact_target() to reconstruct a real (unvalidated) ArtifactGroup/Artifact from the wire-format dict that explainer_job.py actually persists (it normalizes plot()'s output before pickling), so story()'s isinstance checks work against real data, not just hand-built test objects.

DashAI/front/src/components/explainers/ (Frontend)

• StoryBox.jsx: new component that reads the active i18next language from story and renders it as a plain "text" artifact through the existing ArtifactViewer — same box, same download button, nothing new to maintain visually.
• ExplainersPlot.jsx: renders StoryBox under each grouped and ungrouped artifact.
• utils/i18n/locales/{en,es,pt,de,zh}/explainers.json: added the storyTitle label.

Tests

• tests/back/api/test_explainer_story_attach.py (new): covers _attach_stories against the real dict-shaped output explainer_job.py persists (not hand-built Pydantic objects), including the no-explainer no-op case.
• tests/back/explainers/test_image_explainers.py, test_lib_explainers.py, test_new_explainers.py, test_task_explainers.py: updated assertions that expected the now-removed TextArtifact in plot()'s output; assert story()'s content directly instead.

---

Testing

• Manually verified against real generated explanations (not just synthetic test data) for both global and local explainers, in multiple languages, including the flat-curve and zero-importance edge cases.
• Reviewers should spot-check the wording of the non-English translations (es/pt/de/zh) in at least one explainer of each kind (global ranking, local per-instance, counterfactual). They were hand-translated, not machine-translated.

[cristian-tamblay]

Blocking: explanations created before v0.9.7 will 500 on plot load

_attach_stories walks the raw pickle and its normalized counterpart with
zip(..., strict=True), which assumes the pickle is always a list. That only
holds for data written after explainer_job.py started normalizing before
pickling (58c6262dc, 2026-07-10). Tag v0.9.6 is from 2026-06-29, so any
explanation generated on 0.9.5/0.9.6 is still stored in its raw shape: a bare
plotly string or a bare dict.

Running the PR code against those shapes:

raw = bare plotly string   ->  ValueError: zip() argument 2 is shorter than argument 1
raw = bare dict            ->  ValueError: zip() argument 2 is shorter than argument 1
raw = bare dict (local)    ->  KeyError: 0
raw = list (current data)  ->  OK

When raw is a string, zip iterates its characters; when it is a dict, it
iterates its keys. The KeyError: 0 comes from _is_grouped_raw(raw[0]).

Neither call site catches it (explainers.py:491 and :790), so it surfaces
as a 500. These explanations render fine today because normalize_artifacts
handles the legacy shapes, so this is a regression for existing users.

Fix: mirror normalize_artifacts' own wrapping at the top of _attach_stories

if isinstance(raw, (str, dict, Artifact, GroupedArtifacts)):
    raw = [raw]

and wrap the call site in a try/except, so the "never raises" contract in the
docstring actually holds and any future mismatch degrades to "no story" instead
of taking down the plot.

test_explainer_story_attach.py only covers the current normalized shape.
Worth adding the two legacy cases.

[Creylay]

Blocking: explanations created before v0.9.7 will 500 on plot load

_attach_stories walks the raw pickle and its normalized counterpart with zip(..., strict=True), which assumes the pickle is always a list. That only holds for data written after explainer_job.py started normalizing before pickling (58c6262dc, 2026-07-10). Tag v0.9.6 is from 2026-06-29, so any explanation generated on 0.9.5/0.9.6 is still stored in its raw shape: a bare plotly string or a bare dict.

Running the PR code against those shapes:

raw = bare plotly string   ->  ValueError: zip() argument 2 is shorter than argument 1
raw = bare dict            ->  ValueError: zip() argument 2 is shorter than argument 1
raw = bare dict (local)    ->  KeyError: 0
raw = list (current data)  ->  OK

When raw is a string, zip iterates its characters; when it is a dict, it iterates its keys. The KeyError: 0 comes from _is_grouped_raw(raw[0]).

Neither call site catches it (explainers.py:491 and :790), so it surfaces as a 500. These explanations render fine today because normalize_artifacts handles the legacy shapes, so this is a regression for existing users.

Fix: mirror normalize_artifacts' own wrapping at the top of _attach_stories

if isinstance(raw, (str, dict, Artifact, GroupedArtifacts)):
    raw = [raw]

and wrap the call site in a try/except, so the "never raises" contract in the docstring actually holds and any future mismatch degrades to "no story" instead of taking down the plot.

test_explainer_story_attach.py only covers the current normalized shape. Worth adding the two legacy cases.

Fixed by mirroring normalize_artifacts' own wrapping at the top of _attach_stories.

Wrapped both call sites (:491 and :790, now shifted a few lines) in try/except, logging a warning and falling back to "no story" instead of a 500 on any future shape mismatch, so the "never raises" contract in the docstring actually holds.

Added the two legacy-shape regression tests in test_explainer_story_attach.py (bare string for the global path, bare dict for the local/create_grouped=True path). Both reproduce the original ValueError/KeyError against the pre-fix code and pass now.