[G3] Cutover: relayhistory becomes the default and only source; delete the harness readers, walkers, cursors and fs watcher; ship as a major release

[willwashburn]

Part of #553 (group 3 — depends on every burn group-2 issue (#557, #558, #559, #561) and on relayhistory #183's parity report showing 100% identity match for Claude and Codex; parallel with #560/#563).

Scope

1. Default flip: relayhistory-source feature becomes non-optional; BackendChoice::Builtin removed; IngestRoots removed from the public API (lib.rs:132-139 re-exports) — replaced by a thin SessionStoreOptions (db path, provider roots override passthrough for tests) that maps onto ai-hist's StoreOptions. Env: RELAYBURN_SOURCE removed; AI_HIST_DB honoured; BURN_CLAUDE_PROJECTS_DIR removed (was a second root injection point — burn [G1] Remove vestigial ingest surface before the swap: opencode-stream cursor, legacy harnesses/ adapters, stale reader docs #556).
2. Delete (with their tests): crates/relayburn-sdk/src/reader/claude.rs + claude/{incremental,parent_chain,relationships,tool_results,subagents,tests}.rs, reader/codex.rs + codex/{incremental,tests}.rs, reader/opencode.rs + opencode/tests.rs, ingest/{walk,cursors,fs_events,gap,reingest}.rs (gap logic moves to the backend per [G2] Watch mode, --hook claude, pending stamps and gap warnings over the relayhistory backend #558), ingest/backend/builtin.rs, crates/relayburn-cli/src/harnesses/, the notify/memchr deps if unused. Keep: reader/types.rs, reader/{classifier,user_turn,inference,fidelity,git,hash}.rs, reader/claude/span_tree.rs, reader/codex/span_tree.rs (they consume records, not logs — consider renaming reader/ → model/), everything under analyze/, ledger/, ingest/{pending_stamps,watch_loop}.rs as adapted.
3. Fixtures: tests/fixtures/{claude,codex,opencode} stay as the input to the parity/ledger-snapshot suites (relayhistory Add --provider filter to burn compare (closes #138) #163 owns the canonical copy; add a tests/fixtures/README.md saying so and a CI check that burn's copy is byte-identical to relayhistory's tagged copy, or fetch them from the pinned relayhistory tag).
4. Migration for users (docs/migrating-to-5.md + README.md "Quick Start"): npm i -g relayburn now depends on the relayhistory store; first run performs the [G2] Identity and fingerprint compatibility, cursor migration, and state rebuild from the relayhistory store #561 reconciliation; burn ingest = "sync relayhistory then pull"; --hook claude unchanged; burn state rebuild needs the relayhistory DB, not the logs; what happens if the relayhistory DB is missing (bootstrap it: SessionStore::open + sync with default roots — burn must not require the ai-hist CLI to be installed).
5. Docs: README.md first paragraph and burn ingest section; AGENTS.md layout section (drop reader/), "Adding ingest support" replaced by "Session sourcing lives in relayhistory"; CHANGELOG.md [Unreleased] → 5.0.0 narrative: breaking (env vars, IngestRoots, IngestOptions.harness semantics), added (Cursor/Grok via [G3] Harness expansion through relayhistory: Cursor and Grok in burn (SourceKind, pricing, TOOL_ALIASES, overhead mapping, fidelity policy) and collector-backlog transfer #560), fixed (Codex turn accounting collapses many API requests into one turn #504/Attribution and grouping gaps: subagent spend, project rollups, harness split, session token column #506 via [G3] Accounting fixes unlocked by shared evidence: Codex per-request turns (#504), subagent project rollups and harness split (#506), sessions list token column #563).
6. Versioning: workspace version = "5.0.0"; packages/* bumped by the publish workflow; relayburn-sdk crate major bump; pin ai-hist = "=x.y.z" (exact — the crate version is the contract) and add the reverse contract-drift job from relayhistory Wire provider rendering through analyze #183 to ci.yml (parity suite against the pinned published crate). Adopting a new ai-hist version is a deliberate PR that bumps the pin and re-runs parity.
7. Verification before publish: on three real machines, burn summary --since 30d --json before and after upgrade — totals, per-model tokens and cost identical (or differences enumerated in the release notes with root cause). Include the positive control from CLAUDE.md's "verifying claims" discipline in the release checklist.

Acceptance

• grep -rn "parse_claude_session\|parse_codex_session\|parse_opencode_session\|notify::" crates/ → no matches.
• cargo test --workspace green; ledger snapshot suite ([G1] Characterization lock: ledger golden snapshots for the full fixture corpus and a burn state parity ledger differ #555) and parity suite pass with relayhistory as the only source; golden CLI tests unchanged except documented help-text edits.
• Fresh machine: npm i -g relayburn && burn summary works with no relayhistory CLI installed.
• Upgrade machine: no duplicate turns, cost totals identical to pre-upgrade.
• 5.0.0 published; relayhistory Add OpenCode stream wakeups to watch #184's tripwire passes the following week.

Out of scope

New analytics.