feat: unify the recording/forwarding configuration interface.

[JosiahBull]

Summary

Recording storage was configured cluster-wide (ProxyClusterBuilder::storage(...)) while replay was configured per-upstream (a ReplaySource passed to add_upstream_with). The two halves of the same record→replay round-trip were wired through different mechanisms and didn't reference each other — so it was easy to set a replay source per upstream and assume recording would land in the same place, when in fact recording went to a single shared sink (or, with no .storage(...) call, to an in-memory ring that was silently discarded on shutdown).

This PR unifies the two: storage is now configured per upstream, and a single medium drives both directions.

What changed

Public API (breaking)

• Removed ProxyClusterBuilder::storage(...) and the cluster-wide storage field.
• New Snapshots bundle type — the single per-upstream storage knob:
  • Snapshots::from_storage(storage, strategy) — a durable medium (e.g. a JSONL file) used as both the replay source (loaded + indexed at run()) and the recording sink (appended to in Mode::Record).
  • Snapshots::in_memory(exchanges, strategy) — a replay-only source, never written back (tests/fixtures).
• add_upstream_with / add_upstream_with_mode now take Option<Snapshots> instead of Option<ReplaySource>.
• Recorder::with_storage(cfg, Option<SharedStorage>) → Recorder::with_routes(cfg, HashMap<String, SharedStorage>).
• Recorder::storage() → Recorder::storage_for(&str).

Internals

• The recorder stays cluster-wide for the in-memory ring + assertion API (QueryTraffic/AssertSeen/AssertCount are unchanged), but durable storage became a route map keyed by upstream name: record() appends each exchange to its own upstream's medium (durable-before-memory ordering preserved), and flush() fences every backend.
• At run() (already async), each upstream's Snapshots is resolved — its existing contents stream into the indexed ReplaySource, and a durable medium is registered as that upstream's recording sink.

Versioning

• Bumped 0.3.0 → 0.4.0. The API removals/signature changes are breaking; under Cargo's semver for 0.x that's a minor bump, which clears the failing cargo semver-checks gate. Workspace version + internal dep entries + Cargo.lock updated in lock-step.

Docs

• SPECIFICATION.md §3.3, §4, §4.1, §6.4, §8, §9.1, §9.2, §20.1 updated to describe per-upstream Snapshots; removed the dead persist_path / ProxyClusterBuilder::storage references.

Migration

// before
let builder = ProxyClusterBuilder::new()
    .storage(storage)
    .add_upstream_with("api", cfg, mw, Some(replay_source));

// after — one medium per upstream, used for record + replay
let builder = ProxyClusterBuilder::new()
    .add_upstream_with(
        "api", cfg, mw,
        Some(Snapshots::from_storage(storage, MatchStrategy::MethodUriAndBodyHash)),
    );

Behavior note

Because the medium feeds replay too, a re-run in Mode::Record replays any request already present in the file rather than re-recording it (the snapshot acts as a dedup cache, §8.3) and only forwards genuinely new requests. Delete the file to force a clean re-capture.

Testing

• cargo build --workspace --all-targets — clean
• cargo test -p partly-proxy-lib — all suites pass (record 5/5, replay 11/11, lib units 73/73)
• cargo clippy -p partly-proxy-lib --all-targets — clean