Ollama yield is ~1 card per section: cards_per_concept is ignored and there is no retry on zero

[tonserrobo]

Evidence (UE5 C++ Setup run, 2026-07-03)

CLI advertises target: 3 per concept, but the run produced 6 cards from 14 sections (13,432 chars). Nearly every section logs Generated 1 cards or Generated 0 cards; 8/14 sections ended uncovered. The other three runs show the same ~1 card/section ceiling.

Root causes

1. cards_per_concept is threaded through generate_flashcards() but _build_ollama_prompt() never receives it — the prompt just says "Generate 1 to 6 cards" and gemma-class models reliably return exactly 1.
2. Sections are size-capped at 1200 chars and _CHUNK_MAX_CHARS is also 1200, so every section is a single chunk = a single model call = at most ~1 card.
3. When the quality gate or validation drops that one card, the section nets 0 and there is no second attempt — it goes straight to "uncovered".

Fix ideas

• Pass cards_per_concept into the Ollama prompt ("generate ~N cards, one per distinct fact").
• On a 0-card outcome for a non-trivial section, retry once (optionally with a nudge like "the previous attempt produced nothing usable").

Related: #1 (card quality), and the checkpoint issue that freezes these zero results.

[tonserrobo]

Partial fix shipped:

• cards_per_concept now reaches the Ollama prompt ('Aim for about N cards, one per distinct fact').
• A section that nets 0 cards is retried exactly once before going uncovered.

Verified live on UE5 C++ Setup: coverage 6/14 → 8/13, with the retry recovering 2 sections outright. Remaining gap: gemma still mostly returns a single card per call regardless of the stated target, and on 5 stubborn sections that one card fails the quality gate on both attempts (thin/duplicate). Leaving open — the next lever is probably feeding the drop reason back into the retry prompt, or splitting sections below the current 1200-char cap so the model gets one fact per call.