refactor(pedagogy): split prompt_builder.rs into responsibility submodules

[hherb]

Summary

Splits the 708-line primer-pedagogy/src/prompt_builder.rs into a directory module split by responsibility — behaviour-preserving, all files < 500 lines. Tests already lived in the prompt_builder/tests.rs sibling (untouched), so the 708 lines were pure production code.

file	lines	holds
prompt_builder/mod.rs	74	module doc + english_pack() (private, shared) + flat pub use façade
prompt_builder/system_prompt.rs	297	build_system_prompt_* family + assemble_system_prompt core + truncate_passages + days_since
prompt_builder/prompt.rs	151	build_messages + build_prompt_* family
prompt_builder/intent.rs	252	decide_intent* (the Socratic brain) + extract_active_concepts + opener/assertion classifiers

Visibility notes

• english_pack() stays in mod.rs (all three submodules reach it via super::english_pack(); a test child of mod.rs sees it too).
• The build_prompt_* functions call the sibling build_system_prompt_* via the super:: re-exports.
• The (untouched) tests.rs reaches english_pack, the private is_factual_question / is_factual_question_with_pack intent helpers, and the Passage type by their bare names via use super::*. mod.rs carries three #[cfg(test)]-gated use re-imports — matching what the pre-split flat file leaked into module scope — so the test file needs zero edits.

Verification

• Pub-surface diff empty (15 symbols, pub (struct|enum|fn|const|async fn|trait|type), old flat file vs new submodules).
• cargo test -p primer-pedagogy → 216 passed / 0 failed (baseline match).
• cargo clippy --workspace --all-targets -- -D warnings clean.
• cargo fmt --all -- --check clean.
• cargo test --workspace → 51 × test result: ok, 0 failed.

Internal refactor, no runtime behaviour change.

🤖 Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying primer with    Cloudflare Pages

Latest commit:	585dc16
Status:	✅  Deploy successful!
Preview URL:	https://c9fbda27.primer-10b.pages.dev
Branch Preview URL:	https://refactor-split-prompt-builde.primer-10b.pages.dev

View logs

[hherb]

Code review

Found 1 issue:

1. The CLAUDE.md prompt_builder bullet self-cites the wrong PR number: it says "(behaviour-preserving, PR refactor(core): split consts.rs into per-area submodules #322)" but this is PR refactor(pedagogy): split prompt_builder.rs into responsibility submodules #321. Every other reference in this PR's diff (NEXT_SESSION.md, the handoff doc) correctly says refactor(pedagogy): split prompt_builder.rs into responsibility submodules #321, and the sibling split PRs refactor(knowledge): split lib.rs into responsibility submodules #304/refactor(engine): split wiring.rs into responsibility submodules #305 self-cited their own number correctly in this same "(behaviour-preserving, PR #NNN)" pattern. A future bisector following the citation would land on the wrong PR.

primer/CLAUDE.md

Lines 81 to 83 in 7c1c16a

	- **`primer-pedagogy`** — the Primer's "soul". Two modules:
	- `prompt_builder` builds the system prompt (encoding the Socratic method as instructions, varying by age, engagement, and intent), assembles messages from session turns, and contains `decide_intent()` — the heuristic that picks the next `PedagogicalIntent`. `build_prompt` takes a `summary` (rolling pre-window summary) and a slice of `retrieved_older` turns; both are injected as system-prompt sections so the chat-message timeline (the last `context_window_turns` turns) stays linear. **`prompt_builder` is a directory module split by responsibility** (behaviour-preserving, PR #322): `mod.rs` owns the module doc, the process-wide cached `english_pack()` (private, shared by all three submodules via `super::english_pack()`), and a flat `pub use` re-export façade so every `prompt_builder::<name>` path is unchanged (15-symbol pub surface); `system_prompt.rs` the `build_system_prompt_*` family + the shared `assemble_system_prompt` core (base/intent/engagement/summary/retrieved/vocab/knowledge assembly, budgeted + unbudgeted) + `truncate_passages` + `days_since`; `prompt.rs` `build_messages` + the `build_prompt_*` family (which call the sibling `build_system_prompt_*` via the `super::` re-exports); `intent.rs` `decide_intent*` (the Socratic brain) + `extract_active_concepts` + the opener/assertion classifiers (`is_factual_question_with_pack`, `is_confusion_with_pack`, `is_probeable_assertion_with_pack`, `matches_opener`). The pre-existing `prompt_builder/tests.rs` is untouched: it reaches `english_pack` (in `mod.rs`, accessible to the test child) plus the private `is_factual_question`/`is_factual_question_with_pack` intent helpers and the `Passage` type via `use super::*` — `mod.rs` carries three `#[cfg(test)]`-gated `use` re-imports (matching what the pre-split flat file leaked into module scope) so the test file needs no edit.
	- `dialogue_manager` orchestrates a session: record child turn → decide intent → retrieve knowledge → retrieve long-term memory (summary + FTS-relevant older turns) → build prompt → call inference → record Primer turn → refresh summary if due → update learner model. New entry point `resume_session(loaded)` replaces `open_session()` for resumed flows: it swaps in the loaded `Session`, clears `ended_at`, refreshes the summary **only if the loaded one is stale** (covers fewer pre-window turns than the current range; a current summary is preserved verbatim — no point burning an LLM call on identical work), and persists. Two summary-refresh cadences: active-conversation `refresh_summary_if_due` uses a K-threshold (re-summarize when `context_window_turns` new pre-window turns have accumulated, default K=20) so per-turn dialogue doesn't trigger an LLM call every time the boundary advances; resume `refresh_summary_if_stale` uses a simple "is the summary out of date?" check (any uncovered pre-window content triggers a refresh). Summarization goes through `InferenceBackend::summarize` — stub returns a canned string; cloud/ollama default-impl through `generate`.

Otherwise this is a clean, behavior-preserving split: the 15-symbol public surface, english_pack() placement, super:: re-exports, and the #[cfg(test)] re-imports for the untouched tests.rs all check out, and no pedagogical/intent logic changed.

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}