fix(codex): restore decode behaviour lost in the extraction

[ozymandiashh]

Four gaps between the pre-extraction Codex decoder and the one now in @codeburn/core.

Fork-replay filter narrowed from eight event types to one

A forked session replays its parent's history clustered at fork time. The old decoder skipped eight payload types when the timestamp preceded the fork cutoff; this one applied the cutoff only inside the token_count branch.

Token and cost dedup therefore stayed correct — which is exactly why a cost-based parity check cannot see this — while the parent's replayed tool calls, patch applies and MCP tool ends leaked into the child's turn, inflating tool counts, edit LOC and failed-edit counts. Restored to the same position (after turn_context, before dispatch) with the same eight types.

custom_tool_call no longer handled

The old decoder ran function_call and custom_tool_call through one name map; this one handled only function_call, so custom tool calls vanished from a turn's tools and toolSequence. The old custom_tool_call_output branch fed only the task-timing machinery, which does not exist here, so nothing pairs with it.

Buffer path dropped fields the old one synthesised

Lines over 32 KB are scraped field-by-field rather than JSON-parsed, and that path set neither payload.info nor payload.invocation.

Being precise about which matters: info is latent — a token_count record is a few hundred bytes and never becomes a Buffer. invocation is live — an mcp_tool_call_end with large arguments does cross the threshold, and the MCP tool name was being dropped from the turn. Both restored, including the tail-window fallback the old code used and this one had lost.

exec → Bash mapping dropped

Codex Desktop's custom-tool transport uses the short name for the same shell tool.

Verification

Core 513, CLI codex 22. The fork test pins both directions: neutralise the skip and the child wrongly receives the parent's ['Bash','Edit','mcp__srv__t1']; make the skip unconditional and the parent's own assertions fail. A test that only fails one way is half a test.

Two residuals, both inherited rather than introduced

Recorded because they are real and someone should decide about them, not because this PR causes them:

• An invocation sitting in the dead zone — past the 64 KB head, before the 16 KB tail — combined with a complete nested invocation in the tail yields a wrong MCP name rather than a missing one. Inherited verbatim from the old implementation's identical head-then-tail lookup.
• The fork-cutoff string comparison has no NaN guard, so a numeric timestamp silently disables the filter and a malformed one sorting before the cutoff drops legitimate events. Present identically in every version of this code going back through the history — restoring the filter returns the blast radius to its historical baseline rather than widening it. The parse-path guards in fix(codex): validate rollouts structurally, guard the parse path #926 address the adjacent timestamp problems.

[ozymandiashh]

Closing in favour of #940, which now carries these tests.

Cross-PR review established that every source fix here was already present in #940 — the eight-type fork-replay skip, exec to Bash, custom_tool_call dispatch, and the Buffer-path info/invocation synthesis — and that both PRs independently defined the same three raw-field helpers at different positions, so union resolution would have produced three duplicate implementations and one-side resolution would have silently dropped work.

Reducing this PR to its genuine remainder left exactly one thing: the four GAP regression tests. Those turned out to be the part #940 was actually missing — it restored the behaviour without anything pinning it. They now live there, verified failing against a decoder without each fix and green with it.

Keeping this open would have meant a permanently red PR waiting on a sibling, which is not something to hand a maintainer.