[G1] Merge ai-hist-core and ai-hist-engine into one published ai-hist crate with feature gates, a release pipeline and a semver policy #162

Part of #160 (group 1 — independent).

Problem

burn is a Rust workspace (relayburn-sdk is published to crates.io) and can only depend on relayhistory from Rust. relayhistory's Rust crates are workspace-internal and split in a way that helps nobody:

• crates/ai-hist-core/Cargo.toml — ai-hist-core 0.1.0, never released.
• crates/ai-hist/Cargo.toml — package ai-hist-engine 0.2.0 (lib ai_hist_engine), never released.
• Every consumer already depends on both: crates/ai-hist-cli/Cargo.toml:13-14, crates/ai-hist-napi/Cargo.toml:13-14, plugins/provider-sources/rust/Cargo.toml:10-11, plugins/relayhistory/rust/Cargo.toml:15-16. Nothing consumes core alone.
• The layering story is already broken: core holds parse_claude / parse_codex / parse_cursor_text (crates/ai-hist-core/src/lib.rs:1123-1175) and the OpenCode sync (:2124-2222), while the main Claude/Codex parsers live in engine. It is not "storage vs parsing"; it is two halves of one library.
• Public surface is ~250 pub items across the two crates, and 47 functions in core take a raw rusqlite::Connection — an embedder is handed the database driver and a bag of free functions.
• docs/releasing.md describes only the npm pipeline (.github/workflows/publish.yml, name bound to npm OIDC trusted publishing — must not be renamed). There is no cargo publish step anywhere.
• All four candidate names are free on crates.io (ai-hist, relayhistory, ai-hist-core, ai-hist-engine return 404 today).

Both workspaces already agree on rusqlite = 0.32 with bundled, so linking is feasible without a version bump on either side.

Decision

Publish one crate, ai-hist (matching the npm package name), formed by merging crates/ai-hist-core and crates/ai-hist into crates/ai-hist. ai-hist-cli and ai-hist-napi stay as unpublished workspace members depending on it (publish = false). The two plugin Rust crates likewise depend on ai-hist only.

Scope

1. Merge the crates. Move ai-hist-core/src/{storage,observations,relationships,source_evidence,privacy,delivery} under crates/ai-hist/src/ as modules; fold ai-hist-core/src/lib.rs into crates/ai-hist/src/store/ (schema, open/init, migrations, queries). Delete crates/ai-hist-core. Update the four dependents' Cargo.toml. cargo test --workspace must be green with no test body changes beyond use paths — this is a move, not a rewrite.
2. Tighten visibility. Everything that takes a &Connection becomes pub(crate). The only public entry point is the SessionStore facade from [G2] SessionStore facade: the single public entry point (seven methods, typed evidence, no raw connections, no contract constant) #178 (that issue defines the methods; this one makes the rest private so it can). Where the CLI/napi/plugins currently call connection-level functions, route them through the facade or through a #[doc(hidden)] pub mod internal gated behind a unstable-internal feature the workspace enables and external consumers do not. Check with cargo public-api that the default-feature surface is SessionStore, the evidence structs, the error type, and the Source enum only.
3. Feature gates, default off: delivery (the durable delivery coordinator — 37 public items on its own, only the cloud plugin needs it), opencode-backup (the whole-DB Connection::backup path; see [G1] OpenCode: event-level parity across both storage layouts (SQLite opencode.db and legacy JSON storage/), with tokens, model/provider, parts, step-finish, compaction, parentID #168), git-hooks (install_git_hooks / link_git_commit). The napi and CLI crates enable what they need. Verify cargo tree -e features -p ai-hist --no-default-features pulls no tempfile/fs2/url unless a feature asks for it.
4. Make it publishable. description, license (MIT), repository, readme, keywords, categories; workspace rust-version (MSRV) verified with cargo +<msrv> check; cargo publish --dry-run -p ai-hist on CI.
5. Release pipeline. Add a publish-crate job to .github/workflows/publish.yml (do not rename the workflow) after the npm core publish, using crates.io trusted publishing (OIDC) or CARGO_REGISTRY_TOKEN; idempotent when the version already exists (skip, don't fail); honour dry_run.
6. Versioning policy in docs/releasing.md: the crate shares the npm version line (0.18.x today) so one release tag means one artefact set. Cargo semver is the Rust contract — there is no separate Rust contract-version constant. Any change to a public struct field, enum variant, or behaviour a consumer observes is a minor bump pre-1.0 and gets a ### Rust API entry in CHANGELOG.md; a cargo public-api diff in CI (see [G2] Embedder guide, example consumer crate, and a CI job that builds an out-of-tree Rust consumer against the published crates #182) enforces the changelog entry.
7. Docs. docs/releasing.md "Rust crate" section; README.md two-line "Embedding from Rust" pointer; docs/architecture.md call graph updated (one Rust box, not two).

Acceptance

• crates/ai-hist-core no longer exists; cargo test --workspace green.
• cargo public-api -p ai-hist (default features) lists SessionStore, its methods, the evidence structs, Source, Error, and nothing that mentions rusqlite.
• cargo publish --dry-run -p ai-hist passes on CI; a tagged release publishes it; cargo add ai-hist in a scratch crate outside the workspace compiles SessionStore::open(Default::default())?.sync(Default::default())?.
• cargo build -p ai-hist --no-default-features compiles without the delivery module.

Out of scope

The facade method set (#178); any parser change.