Merge #1718: feat(tracker-core): migrate SQLite and MySQL drivers to sqlx 0.8

66f1246 test(tracker-core): wait for downloads==1 instead of asserting on intermediate state (Jose Celano)
fd5be6d fix(tracker-core): accept no-op SQLite upserts as success (Jose Celano)
c1cce75 docs(bootstrap): restore #[must_use] and fix typo on setup() (Jose Celano)
78c63a4 test(tracker-core): bound persistence retry with timeout (Jose Celano)
53297d3 fix(tracker-core): build SQLite connection options from filesystem path (Jose Celano)
61e4427 fix(tracker-core): reject negative valid_until and propagate row-count overflow (Jose Celano)
2c19024 fix(tracker-core): drop torrent_aggregate_metrics table on schema teardown (Jose Celano)
5512ac3 fix(tracker-core): accept no-op MySQL upserts as success (Jose Celano)
0990113 docs(agents): clarify how Committer handles split commits with cspell changes (Jose Celano)
6510da5 docs(tracker-core): record 2026-04-30 persistence benchmark run (Jose Celano)
42e3e5d test(tracker-core): wait for mysql readiness in persistence benchmark (Jose Celano)
a4dbc63 refactor(tracker-core): remove legacy r2d2 persistence surface and temporary sqlx staging tree (Jose Celano)
8c07450 docs(1525-05): track remaining cleanup and validation work (Jose Celano)
93e25f3 refactor: propagate async initialization end-to-end removing all run_on_runtime bridges (Jose Celano)
18b3b0f refactor(tracker-core): complete async sqlx switch and add task 5 cleanup (Jose Celano)
aff5bd2 refactor(tracker-core): remove lazy schema checks from sqlx drivers (Jose Celano)
ed0cef1 docs(issues): keep eager schema initialization in 1525-05 (Jose Celano)
76594b7 feat(tracker-core): add async sqlx mysql driver in parallel module (Jose Celano)
2fb25a1 docs(skills): require opening PRs in upstream repo (Jose Celano)
8cc5deb feat(tracker-core): add async sqlx sqlite driver in parallel module (Jose Celano)
91547d5 feat(tracker-core): add sqlx 0.8 infrastructure and async error conversions (Jose Celano)
333f6ef docs(issues): correct 1717 spec to match current codebase structure (Jose Celano)
8f10e39 docs(issues): rename 1525-05 spec to include GitHub issue number 1717 (Jose Celano)

Pull request description:

  Closes #1717

  Migrates the SQLite and MySQL persistence drivers from the sync r2d2 stack to async sqlx 0.8.

  ## Current status

  Structural migration and post-migration benchmark validation are complete.

  Completed on this branch:

  - SQLite and MySQL driver implementations use `sqlx` pools and async trait methods.
  - Schema initialization remains eager via `initialize_database()`.
  - Schema creation still uses raw `sqlx::query()` DDL; `sqlx::migrate!()` is not introduced here.
  - Sync-to-async runtime bridge helpers have been removed; async initialization is propagated through all current call paths.
  - The temporary `packages/tracker-core/src/databases/sqlx/` staging tree has been deleted; the canonical `databases/driver/` and `databases/traits/` directories are now the single persistence surface.
  - Legacy dependencies (`r2d2`, `r2d2_sqlite`, `r2d2_mysql`) have been removed from `packages/tracker-core/Cargo.toml`.
  - Legacy compatibility code has been removed: the `Error::ConnectionPool` variant and all `From` impls for `r2d2::Error`, `rusqlite::Error`, `mysql::Error`, and `mysql::UrlError` are gone from `databases/error.rs`. The `From<rusqlite::Error>` impl in `authentication/key/mod.rs` was replaced with `From<sqlx::Error>`.
  - Stale `r2d2_*` references in driver doc comments have been refreshed to describe the sqlx-based pools.
  - Validation passing: `cargo machete`, `linter all`, doc tests, and `cargo test --workspace --all-targets`.
  - Persistence benchmarks re-run and compared against the 2026-04-28 baseline. Report: [`packages/tracker-core/docs/benchmarking/runs/2026-04-30/REPORT.md`](packages/tracker-core/docs/benchmarking/runs/2026-04-30/REPORT.md). No regression detected: MySQL totals are ~13–16% faster (mysql 8.4: 7381 → 6231 ms; mysql 8.0: 7633 → 6678 ms); SQLite totals shift within expected per-run jitter on the 100-op suite.

  See `docs/issues/1717-1525-05-migrate-sqlite-and-mysql-to-sqlx.md` for the full task breakdown and progress review.

  ## 2026-04-30 update

  - **Task 7 (persistence benchmarking) is done.** See the report linked above.
  - **Bench harness change** (`42e3e5df`): the `persistence_benchmark` MySQL setup now waits for the second `ready for connections` line on the container's stderr (the TCP listener, not the local socket) and then pings `SELECT 1` in a 30 × 500 ms retry loop before initializing the schema. This is needed because `sqlx` does not retry the first `connect()` (unlike the old r2d2 pool), so the harness must guarantee MySQL is accepting TCP connections before the benchmark proceeds.
  - **Production-code follow-up** (`6510da53` / related): expanded the `# Panics` Rustdoc section of `bittorrent_tracker_core::databases::setup::initialize_database` to document that it will panic if the database is not yet accepting connections, or on any other underlying `sqlx::Error`. This makes the new `sqlx`-based connect-without-retry behavior explicit to callers.

  ## 2026-04-30 update — Copilot review addressed

  Copilot performed an automated review on this PR (15 inline comments). All actionable items have been resolved across 6 GPG-signed commits (`5512ac30..c1cce75`); 1 item was already obsolete.

  | Commit | Subject | Addresses |
  |---|---|---|
  | `5512ac30` | fix(tracker-core): accept no-op MySQL upserts as success | #3, #4 |
  | `2c190249` | fix(tracker-core): drop torrent_aggregate_metrics table on schema teardown | #5, #6 |
  | `61e44277` | fix(tracker-core): reject negative valid_until and propagate row-count overflow | #8, #11, #12, #13, #14, #15 |
  | `53297d3e` | fix(tracker-core): build SQLite connection options from filesystem path | #10 |
  | `78c63a4d` | test(tracker-core): bound persistence retry with timeout | #1 |
  | `c1cce750` | docs(bootstrap): restore #[must_use] and fix typo on setup() | #2, #9 |

  Highlights:

  - **MySQL upsert semantics**: removed the `rows_affected() == 0` check on `INSERT … ON DUPLICATE KEY UPDATE` paths — a no-op update legitimately reports 0 affected rows and was being misclassified as `InsertFailed`.
  - **Data integrity**: negative `valid_until` timestamps now return `Error::MalformedDatabaseRecord` instead of being silently flipped positive via `unsigned_abs()`. Row-count → `usize` conversions now propagate errors instead of `unwrap_or(0)`.
  - **SQLite path handling**: switched from `SqliteConnectOptions::from_str("sqlite://{path}")` to `SqliteConnectOptions::new().filename(path)` so relative paths like `./storage/...` aren't reinterpreted as URL authority.
  - **Schema teardown**: `drop_database_tables()` in both drivers now drops the `torrent_aggregate_metrics` table that was being created but never cleaned up.
  - **Test determinism**: replaced the fixed-iteration `yield_now()` retry loop in `it_should_persist_the_number_of_completed_peers_for_each_torrent_into_the_database` with `tokio::time::timeout(5s, ...)` + `sleep(50ms)`.

  Item #7 (sqlx subdir not wired in) was already resolved in commit `a4dbc63a` earlier in this PR.

  Validation: `linter all` exit 0, full workspace test suite passing, MySQL-gated tests passing.

ACKs for top commit:
  josecelano:
    ACK 66f1246

Tree-SHA512: 3c13001f502ecc00579896fe7297981241b5632e263eeb913e542658ec581de28d3ccd2ecf056629e817fd39bb6aa5d36d453b9d0bb2998fabd5a608bb17c47b

Author: josecelano

.github/agents/committer.agent.md

@@ -41,6 +41,19 @@ Treat every commit request as a review-and-verify workflow, not as a blind reque
 - Do not mix skill/workflow documentation changes with implementation changes — always create
   separate commits.
 
+## Splitting Commits
+
+When the requested work spans multiple logical commits and `project-words.txt` has been
+modified with new entries that belong to different commits, do not try to split the
+dictionary additions across those commits. Instead:
+
+1. Commit all `project-words.txt` changes first as a single `chore(cspell): add <words>`
+   commit (or fold them into the first logical commit when that is more natural).
+2. Then create the remaining focused commits for the actual implementation/docs changes.
+
+This keeps the spell-check linter green at every commit and keeps the substantive commits
+focused on their real intent rather than on dictionary churn.
+
 ## Output Format
 
 When handling a commit task, respond in this order:

.github/skills/dev/git-workflow/open-pull-request/SKILL.md

@@ -18,10 +18,14 @@ metadata:
 Before opening a PR:
 
 - [ ] Working tree is clean (`git status`)
+- [ ] Upstream target repository confirmed from workspace metadata (`Cargo.toml` → `repository`)
 - [ ] Branch is pushed to your fork remote
 - [ ] Commits are GPG signed (`git log --show-signature -n 1`)
 - [ ] All pre-commit checks passed (`linter all`, `cargo machete`, tests)
 
+> Important: always open the PR in the **upstream repository**, not in your fork.
+> Resolve upstream from `Cargo.toml` (`repository = "https://github.com/torrust/torrust-tracker"`) and use that value for `gh pr create --repo ...`.
+
 ## Title and Description Convention
 
 PR title: use Conventional Commit style, include issue reference.
@@ -42,13 +46,20 @@ PR body must include:
 
 ```bash
 gh pr create \
-  --repo torrust/torrust-tracker \
+  --repo <upstream-owner>/<upstream-repo> \
   --base develop \
   --head <fork-owner>:<branch-name> \
   --title "<title>" \
   --body "<body>"
 ```
 
+Example upstream resolution from `Cargo.toml`:
+
+```bash
+UPSTREAM_REPO=$(grep '^repository\s*=\s*"https://github.com/' Cargo.toml | sed -E 's#.*github.com/([^\"]+).*#\1#')
+gh pr create --repo "$UPSTREAM_REPO" --base develop --head <fork-owner>:<branch-name> --title "<title>" --body "<body>"
+```
+
 If successful, `gh` prints the PR URL.
 
 ## Option B: GitHub MCP Tools