prebake: always re-ingest so contextual retrieval runs at premium

The previous skip-if-already-indexed short-circuit in scripts/prebake_repos.py
meant contextual retrieval never re-ran for repos that were already in
Qdrant. Result: tour/diagram/README artifacts were premium-quality, but
the underlying chunk vectors kept their original (free-tier or none)
contextual descriptions. Chat retrieval — which depends on contextualised
chunks — silently stayed on the lower tier.

The fix is one line: don't short-circuit. Always pass force=True. The
ingestion service's force=True path runs contextual retrieval enrichment
on every chunk; with premium_mode on, those calls route to claude-sonnet.

Voyage embeddings are deduplicated by content hash, so re-ingestion only
re-embeds changed chunks. Net cost is dominated by the per-chunk
contextual LLM call (premium tier).

Visible effect: every prebaked repo gets the "Contextual retrieval
applied" sparkle in the sidebar afterward. CLAUDE.md updated to reflect
the new behaviour and document the save-site protection from the
previous commit.

CLAUDE.md

@@ -118,6 +118,10 @@ The script flips `gen.premium_mode = True` for the entire run, which:
 - Routes every `gen.generate(...)` call to the Claude Sonnet 4.6 client (`ANTHROPIC_API_KEY` required).
 - Activates `PREMIUM_CAPS` overrides in `GenerationService` — every `gen.cap(name, default)` call returns the larger premium value (longer ReAct rounds, fuller chunk previews in contextual retrieval, larger README budget, etc.).
 
+The script always force-re-ingests each repo, even when already indexed, so contextual retrieval re-runs with the premium model. Without this, chat retrieval stays on free-tier (or non-existent) contextual descriptions. The "Contextual retrieval applied" sparkle in the sidebar is the visible proof that this ran.
+
+**Save sites refuse to demote premium artifacts.** When the runtime UI's Regenerate button fires a free-tier generation, `_save_tour` / `_save_diagram` / the README cache write detect the existing payload's `generated_by_model` starts with `claude-` and skip the persist with a `[protect] not overwriting premium` log line. The user's session shows their regenerated content (in-memory cache updates) but the durable cache stays at premium quality.
+
 Runtime requests from the deployed app keep the original (smaller) caps so free-tier providers don't drown.
 
 To inspect what's been baked for a repo:

scripts/prebake_repos.py

@@ -66,17 +66,26 @@ def repo_indexed(store: QdrantStore, repo: str) -> bool:
 
 
 def ingest(repo: str, store: QdrantStore, gen: GenerationService, embedder: Embedder) -> bool:
-    """Ingest a repo via GitHub. Skips if already indexed."""
-    if repo_indexed(store, repo):
-        print(f"  ✓ already indexed ({store.count(repo=repo)} chunks) — skipping ingestion")
-        return True
-    print(f"  ▸ ingesting {repo}…")
+    """Re-ingest a repo via GitHub with force=True so contextual retrieval
+    runs. Even when the repo is already indexed we re-run — premium prebake
+    must end with premium-quality contextual descriptions on every chunk,
+    not just whatever the previous (possibly free-tier) ingestion left
+    behind. The Voyage embeddings are deduplicated by content hash so this
+    isn't as expensive as it sounds: only changed/new chunks pay the
+    embed cost; only chunks needing fresh contextual retrieval pay the
+    LLM cost."""
+    already = repo_indexed(store, repo)
+    if already:
+        print(f"  ▸ re-ingesting {repo} ({store.count(repo=repo)} chunks already indexed)…")
+    else:
+        print(f"  ▸ ingesting {repo}…")
     ingestion = IngestionService(store=store, embedder=embedder, gen=gen)
     repo_url = f"https://github.com/{repo}"
     try:
-        # force=True so contextual retrieval enrichment runs (premium_mode
-        # is on, so the per-chunk descriptions are generated by the premium
-        # model). progress callback prints sparse milestones to stdout.
+        # force=True triggers contextual retrieval enrichment. Because
+        # premium_mode is on, gen.generate() routes those calls to the
+        # premium client → claude-sonnet-4-6. progress callback prints
+        # sparse milestones to stdout for visibility.
         last_step = [""]
         def on_progress(step: str, detail: str) -> None:
             if step != last_step[0]: