Fork chats from selected messages

[0xSolarPunk]

Summary:

• Add per-message fork actions that send an optional upToSeq cutoff through the chat API and session controller.
• Preserve native JSONL source metadata so forks can truncate at a clicked Claude/Codex message, preferring Claude-style entry ids and falling back to source lines.
• Allow message-point forks while the source is running when the agent supports fork-while-running, and truncate before active/corrupt JSONL tails.
• Pass raw copied Codex rollout paths through resume params and harden HTTP compression/range handling for the current Bun runtime.

API note:

• POST /api/v1/chats/fork now accepts optional upToSeq; omitting it preserves existing full-chat fork behavior.

Dogfood evidence:

• Isolated server from this branch with GARCON_DISABLE_AUTH=true, separate GARCON_CONFIG_DIR, and a seeded Claude JSONL chat containing four complete lines plus one deliberately incomplete active tail.
• Flow tested in browser: open seeded chat, open the first message three-dot menu, click Fork, verify new forked chat opens and renders only the first message.
• Native verification: forked JSONL line count is 1, contains the first prompt, excludes the second prompt, excludes tail response, excludes incomplete active tail, and parses as valid JSONL.
• Report: docs/pr-evidence/message-point-fork/report.md

Screenshots:

Source chat with completed messages and incomplete tail ignored:

First message menu showing per-message Fork:

Forked chat containing only the selected prefix:

Validation:

• bun run check
• bun run test
• timeout 60s bun run start --port 0 || test $? -eq 124 (server reached http://0.0.0.0:34091 before the timeout shut down the smoke process)