djust 0.9.1

• RichSelect variant support, trigger tinting, and onclick parity (#1204). Each option dict accepts an optional variant key that tints the row in the dropdown AND the trigger when that option is currently selected. Built-in variants align with Badge/Button/Tag/Alert vocabulary (info, success, warning, danger, muted, primary, secondary). New variant_map kwarg mirrors Badge.status() for value→variant mapping cases. Permissive variant-name regex (^[a-z0-9][a-z0-9-]{0,31}$) lets downstream projects ship custom variants by adding a matching .rich-select-option--variant-<name> CSS rule. Trigger now emits the open/close onclick handlers that {% rich_select %} template tag always emitted, eliminating the monkey-patch-rendered-HTML workaround programmatic consumers used to need.

• LiveViewTestClient.render_with_patches() — VDOM-diff accessor for tests (#1208). New public method on djust.testing.LiveViewTestClient that wraps view_instance.render_with_diff() and returns (html, patches_list, version) with the JSON patches parsed into a Python list. Empty list when no patches were produced. Reusable for any test that needs to assert on VDOM-diff invariants (e.g. "this noop event must produce zero patches"). First user is the strengthened test_normalize_idempotent_on_already_serialized regression test in tests/unit/test_list_model_diff_1205.py, which now locks the #1206 normalize-pass idempotency contract via an explicit patches == [] assertion instead of the prior weaker "no exception" check.

• JIT serializer silently degrades when context value is list[Model] (#1205, expanded by #1207). When a view's get_context_data override sets ctx["tasks"] = list(qs) after calling super().get_context_data(), the JIT auto-serialization pipeline runs inside super() and never sees the user-added value. The raw list[Model] then flows through _sync_state_to_rust, where change-detection compares list elements via Python == — which delegates to Model.__eq__ (pk-only). In-place field mutations (is_active toggle, completed flip) don't change pk, so the comparison returns equal, the key is never added to the diff-context, Rust never receives the new state, and the rendered HTML is byte-identical on every event despite confirmed DB writes. Symptom from the issue reporter: patch_count: 0 on every event, _debug.variables.tasks.value shows only __str__ strings.

  Initial fix (#1206): _sync_state_to_rust (in python/djust/mixins/rust_bridge.py) now runs a defensive normalize pass over full_context immediately after fetching it, converting any homogeneous list[Model] / Model / QuerySet value to dicts via normalize_django_value. After normalization, change-detection compares list[dict] != list[dict] element-wise via dict.__eq__ (structural), correctly catching field mutations. Idempotent on already-serialized values. Also removed dead _lazy_serialize_context method from python/djust/mixins/jit.py (zero call sites — was misleadingly cited as the bug location in the issue).

  Shape coverage expansion (#1207): the initial fix only handled homogeneous list[Model]; PR-review surfaced two more shapes that escape change-detection — heterogeneous [dict, Model] (Model not first; is_model_list checked only value[0]) and nested list[list[Model]] (grouped tasks). Refactored the inline normalize loop into a recursive _normalize_db_values helper that scans the full list for any-position Model and recurses into nested lists with bounded depth (_NORMALIZE_DEPTH_LIMIT = 3). 9 regression cases in tests/unit/test_list_model_diff_1205.py lock down all shape variants: homogeneous list[Model], single Model, raw QuerySet, heterogeneous [dict, Model], nested list[list[Model]], idempotency, empty list, and mixed-type list.

• Broadcast renders now refresh _recovery_html (#1202). server_push in python/djust/websocket.py was sending broadcast patches without updating self._recovery_html / self._recovery_version (the user-initiated event path did this; broadcasts were overlooked). Consequence: when a subsequent applyPatches on the client failed (the well-known {% if %}-shifts-DOM case), the client sent request_html expecting fresh HTML, but handle_request_html read a stale or None-valued _recovery_html and returned recoverable: false. The client then triggered a full-page reload. Sessions that only received broadcasts after mount (admin dashboards, real-time apps with Celery pushes) never populated _recovery_html via the normal path. PR #1203 mirrors the handle_event recovery-state-store at websocket.py:3271: before sending broadcast patches, server_push now stores the render output as _recovery_html / _recovery_version. Two regression cases in tests/unit/test_server_push.py lock the store-after-broadcast invariant + the no-patches branch.

The v0.9.1 arc shipped a substantial body of internal-tooling work that doesn't change user-facing API but improves contributor and maintainer ergonomics. Highlights for the audit trail (no migration needed):

• Pre-push lints: scripts/check-no-dead-private-methods.py (#1209), scripts/check-no-comma-list-closes.py (#1227), scripts/audit-pipeline-bypass.py (#1212).
• Pipeline-template Stage 4 reproducer-first mandatory item (#1210), Stage 11 reviewer-prompt budget guidelines (#1211), and the two-commit shape canonicalization + 3-clean-runs verification (Action Tracker #181/#182) as structural gates.
• CodeQL sanitizer model for djust.security.log_sanitizer.sanitize_for_log (#1214) — closes the FP class PR #1201's 8 dismissals worked around.
• CLAUDE.md "Bug-report triage" section citing PR #1206 as the canonical case study for issue-reporter-analysis-not-equal-root-cause discipline (#1213).
• 20+ retro patterns canonicalized across CLAUDE.md sections for v0.6.x–v0.8.x retro arcs (#1226), v0.9.4 retro arc (#1225).
• Pre-commit .pxd exclude prevents binary archive corruption (#1215).
• Hot-reload auto-enable via DjustConfig.ready() in DEBUG mode (#1190).
• Debug Panel UI for time-travel + forward-replay (#1151 / PR #1194, on top of v0.9.0's wire-protocol foundation).