[G1] Tool-result fidelity: output bytes, truncation markers, content hash, ordering indexes, out-of-band error status, subagent-notification linkage #171

Part of #160 (group 1 — independent).

Problem

burn's hotspots, tool_output_bloat, context_delta, and span-tree features are driven by per-tool-result facts that relayhistory does not record. relayhistory stores a tool result as a session_events row (role='tool_result', kind='tool_result', text=<content>) and sets tool_calls.is_error (crates/ai-hist/src/lib.rs Claude :3142-3298, Codex :2101-2508). Missing, with burn's reader/consumer for each (crates/relayburn-sdk/src/):

Also: relayhistory's Claude is_error comes only from the tool_result block; Codex's from exit_code/patch_apply_end.success/mcp_tool_call_end.result.Err — good, but the event row does not say which signal set it.

Scope

1. Schema: add to session_events (tool_result rows) payload_bytes INTEGER, payload_truncated INTEGER, payload_hash TEXT, call_index INTEGER, event_index INTEGER, result_status TEXT, event_source TEXT, error_signal TEXT (tool_result.is_error | exit_code | patch_apply | mcp_err | subagent_status), subagent_session_id TEXT, agent_id TEXT. Migration marker + HYDRATION_PARSER_VERSION bump.
2. Claude: compute bytes/hash/truncation over the raw content (string or array, stable-stringified with sorted keys to match burn's stable_stringify), before the text column is filled; maintain per-tool_use_id call_index and a transcript-global event_index counter (persist the last index in the hydration checkpoint so incremental hydration continues the sequence — coordinate with the incremental-hydration issue).
3. Codex: buffer results per turn and back-patch error_signal/result_status at task_complete exactly as burn does; function_call_output rows get event_source='function_call_output'.
4. OpenCode/Grok/Cursor: same columns from their parsers (lands with each parity issue; this issue provides the helper ToolResultFacts::from_payload(&Value) in core they all call).
5. Per-user-row blocks: expose session_user_turns_page(source, session_id) returning, per user message, the ordered [{kind, tool_use_id, byte_len, is_error}] — derivable from session_events (no new table) but computed once in core so every consumer agrees. approx_tokens is not computed here (burn's heuristic; a real tokenizer can be added later).
6. SessionEvent struct, napi, TS contracts, docs/session-catalog.md matrix.

Acceptance

• Corpus oversized-bash-output: the tool_result event has payload_bytes equal to burn's value for the same fixture (assert the number, taken from burn's reader/claude/tests.rs), payload_truncated = 1.
• replacement-meta: event_source and marker linkage present.
• Codex oversized-shell-output + with-tool-call: error_signal = 'exit_code' on the failing call; result_status = 'errored'.
• system-subagent-notification: result row with event_source='subagent_notification', subagent_session_id set.
• session_user_turns_page on user-turn-blocks (Claude, Codex, OpenCode) returns the same block sequence burn's UserTurnRecord tests assert.

Out of scope

Token estimation, usage attribution to tool calls, cost.