[G1] Remove vestigial ingest surface before the swap: opencode-stream cursor, legacy harnesses/ adapters, stale reader docs

[willwashburn]

Part of #553 (group 1 — independent, small). Shrinks what the group-2 backend has to reproduce.

Findings

• crates/relayburn-cli/src/commands/ingest.rs:40 says "The OpencodeStreamIngestor parser still lives in relayburn-sdk::reader::opencode_stream". That module does not exist. The opencode-stream cursor variant in crates/relayburn-sdk/src/ingest/cursors.rs:80-81 is kept only for round-trip compatibility of old archive_state.upstream_cursors_json blobs (cursors.rs:93-118 preserves unknown variants verbatim anyway).
• crates/relayburn-cli/src/harnesses/ (claude.rs, codex.rs, opencode.rs, pending_stamp.rs, registry.rs, test_env.rs) is documented as legacy/reference: the CLI no longer exposes a launch command (harnesses/mod.rs:1-7); only claude is registered (registry.rs:68-70); RUNTIME_ADAPTERS for codex/opencode is empty. The only live use is session_store_adapter for pending-stamp resolution.
• AGENTS.md "Adding ingest support" describes adding readers to crates/relayburn-sdk/src/reader/ and roots to IngestRoots — the path that is being retired.
• BURN_CLAUDE_PROJECTS_DIR is read directly in query_verbs/summary/compute.rs:223 and query_verbs/flow.rs:281 for subagent counts, bypassing IngestRoots — a second, undocumented root injection point.
• reader/user_turn.rs:128-138 errors when a cl100k tokenizer is requested ([Rust port] relayburn-sdk: embedding API mirroring TS surface #246); every approx_tokens is ceil(bytes/4).

Scope

1. Delete the opencode-stream mention from the CLI doc comment; keep the cursor variant (compat) but mark it #[doc(hidden)] with a comment saying it is preserved for old blobs only and will be dropped with the cursor format at cutover.
2. Collapse crates/relayburn-cli/src/harnesses/ to the one function still used by pending-stamp resolution, or move that function next to ingest/pending_stamps.rs in the SDK and delete the directory; update crates/relayburn-cli/tests/smoke.rs harness-name expectations.
3. Route the two BURN_CLAUDE_PROJECTS_DIR reads through the same roots object ingest uses (or through the backend seam from the sibling group-1 issue once it lands — whichever merges first; the other rebases).
4. AGENTS.md: replace "Adding ingest support" with a short "Session sourcing" section: readers are being retired; new harnesses are added in AgentWorkforce/relayhistory; burn adds only pricing, TOOL_ALIASES, overhead-file mapping and fidelity policy per source. Link the burn epic and the relayhistory epic.
5. Document the cl100k gap as a burn-side policy decision in the same section ("byte/4 heuristic; a tokenizer would live in burn, not the sourcing layer").

Acceptance

• grep -rn opencode_stream crates/ returns only the hidden cursor variant.
• crates/relayburn-cli/src/harnesses/ is gone or reduced to one file with a one-line doc comment; cargo test --workspace green; golden tests unchanged.
• AGENTS.md no longer instructs contributors to add readers to burn.

Out of scope

Deleting the readers themselves (group 3).