fix(claude): .orphaned-* segments no longer vanish from chained conversations

[bdelanghe]

Defect 2 from the #236 root-cause analysis (#236 (comment)).

ChainIndex::refresh classified any stem whose first entry carried a foreign sessionId as a chain successor. Rotation artifacts (<uuid>.orphaned-<ts>-<hash>) always do — their entries keep the original session's id while the stem is mangled — so they were chained in, whereupon is_bridge_entry matched every entry of the segment and the merge dropped the whole thing: turns and token_usage silently gone, while claude_chain_stamp still summed the segment's bytes into the freshness stamp. A session stem never contains a dot; dotted stems now stand alone as their own conversations.

Second, read_conversation applied the bridge filter to every entry of every segment, though bridge entries are only ever the predecessor's tail copied to the top of a successor file (as chain.rs's own docs say). The filter now skips only the leading bridge run — a foreign sessionId deeper in a segment is data, kept.

toolpath-claude 0.13.0 → 0.13.1 (crate + workspace Cargo.toml, site/_data/crates.json, CHANGELOG).

Run sheet

1. Independent PR — no bundled or speculative changes
   • Verified: chain.rs classification + lib.rs bridge filter only; the agent-coding-session: per-group token_usage is last-call-wins — token sums undercount and depend on sync granularity #236 sync-merge fix (defect 1) and path-id truncation (defect 3) are deliberately separate work
2. Changed codepaths verified — targeted unit and full integration tests
   • test_orphaned_stem_is_not_a_successor (new), test_bridge_skip_is_leading_run_only (new), full cargo test -p toolpath-claude — 259 green
   • path-cli integration: 42 passed / 20 failed identically on unmodified main in the same local nix-shell env — pre-existing environment failures, not regressions; this CI run is the arbiter
3. Root cause identified — every failure traced to source
   • Orphaned stems misread as successors (chain.rs:94-101) + bridge filter applied beyond the leading run (lib.rs:156-161); live repro: 50357658-….orphaned-1787626221622-ac84712d.jsonl, 8 entries, all silently dropped
4. No duplication — refactoring preferred over copy/paste
   • N/A — filter logic tightened in place
5. No unrelated changes — housekeeping isolated to its own branch
   • Version-bump files are the required release checklist for the fix itself

🤖 Generated with Claude Code

https://claude.ai/code/session_01EbQMJRxBqTQyP7nE7qqzsr

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

🔍 Preview deployed: https://eb632988.toolpath.pages.dev

[bdelanghe]

Amended after checking the fix against the kind spec (toolpath.net/kinds/agent-coding-session/v1.1.0): as first pushed, standalone orphans would have become listed conversations, and derive_path truncates the view id to 8 chars — for <uuid>.orphaned-* that's the parent session's prefix, so the orphan doc and the full session doc would have overwritten each other in the cache (a defect-3 interaction the spec's Σ-guarantee makes obvious). Dotted stems are now excluded from chain_heads() too: neither chained (the data-loss bug) nor listed (the collision). read_segment reads them directly; ingesting them lands with the full-stem id work. New assertion added to test_orphaned_stem_is_not_a_successor.