pika-git-legacy-pr-1

A single line is added to main.rs to register the new module:

mod legacy_pr_store;

This is placed alphabetically between inbox_store and live, following the existing convention. No other changes to main.rs are required because all public symbols are re-exported through storage.rs.

The new file begins with #![allow(dead_code)] at the module level since several structs and methods are not yet called from production code paths. It imports crate::storage::Store so it can add impl Store blocks—Rust allows trait-free impl blocks on a type from another module within the same crate.

All six data structs are moved without modification:

The moved methods and their responsibilities:

1. upsert_pull_request — Inserts or updates a PR row and its associated artifact version within an Immediate transaction.
2. update_poll_markers — Updates repo-level polling timestamps. The implementation is rewritten: instead of upserting into a poll_markers table, it now updates last_open_polled_at and last_merged_polled_at columns directly on the repos table.
3. claim_pending_generation_jobs — Selects pending artifact versions, atomically sets them to generating, and returns GenerationJob records. The new version uses transaction_with_behavior(Immediate) instead of unchecked_transaction and adds a pr.state IN ('open', 'merged') filter to avoid claiming jobs for closed PRs.
4. mark_generation_ready — Completes a generation by storing tutorial JSON/HTML/diff, handling the superseded-vs-activated logic.
5. mark_generation_failed — Records failure with retry-backoff logic (up to 5 retries).
6. close_stale_open_prs — Marks PRs as closed if they no longer appear in the current open set.
7. list_feed_items — Queries the activity feed with latest artifact status.
8. list_pr_summaries — Dynamically-filtered query returning PR summaries with optional tutorial JSON.
9. get_pr_detail — Full detail lookup for a single PR.
10. recover_stale_generating — Startup recovery that resets generating artifacts back to pending.
11. queue_regeneration — Creates a new pending artifact version for re-generation.

All 529 lines of method bodies and 162 lines of helper functions are removed from storage.rs.

Six private (crate-local) functions are placed at the bottom of the new file:

• ensure_repo — Upserts a repos row and returns its id.
• upsert_pr_row — Handles the insert-or-update logic for pull_requests, returning the row id, previous head SHA, and whether an insert occurred.
• upsert_artifact_row — Decides whether to create a new pending artifact version or leave the existing one alone based on head SHA matching and current status.
• insert_pending_artifact_version — Computes the next version number and inserts a pending artifact row.
• supersede_non_current_generations — Marks all non-current pending/generating/failed artifacts as superseded.
• has_inflight_generation_for_head — Checks if a pending or generating artifact already exists for a given head SHA.

These remain private to the crate and are only called by the impl Store methods in the same file.

The import section of storage.rs is cleaned up:

• chrono::Utc is removed (no longer used here).
• TransactionBehavior is removed from the rusqlite import.
• A pub use re-export block is added:

#[allow(unused_imports)]
pub use crate::legacy_pr_store::{
    FeedItem, GenerationJob, PrDetailRecord, PrSummaryRecord,
    PrUpsertInput, UpsertOutcome,
};

This ensures any module that previously imported these types via crate::storage::PrUpsertInput (etc.) continues to work. The #[allow(unused_imports)] suppresses warnings for types that may not currently be referenced through this re-export path.

The Store struct definition and its core infrastructure methods (open, with_connection) remain in storage.rs alongside all non-PR-related logic (inbox methods, chat allowlist, etc.).

While this branch is primarily a code-move refactor, three behavioral changes are introduced:

1. Transaction isolation for claim_pending_generation_jobs: Changed from unchecked_transaction() to transaction_with_behavior(TransactionBehavior::Immediate). This acquires a write lock immediately rather than deferring it, preventing potential SQLITE_BUSY errors when the UPDATE runs inside the transaction. The claim loop is also restructured to first collect candidate IDs, then claim-and-load each one individually with a lost-update check (AND status = 'pending').

2. PR state filter on pending jobs: The generation jobs query now includes AND pr.state IN ('open', 'merged'), preventing artifact generation from being claimed for PRs that have been closed. Previously, a closed PR with a pending artifact would still be picked up.

3. Poll markers storage model: update_poll_markers no longer upserts into a poll_markers table. Instead, it updates last_open_polled_at and last_merged_polled_at columns directly on the repos table. The method signature also changes from (open_cursor, merged_cursor) to (last_open_polled_at, last_merged_polled_at), reflecting the shift from opaque cursor strings to timestamp-based polling.