fix(replay): walk full timeline ancestry in get_full_timeline_steps/spans

[risjai]

Summary

• get_full_timeline_steps and get_full_timeline_spans now walk the entire ancestor chain via an iterative ancestry_chain helper that applies a cumulative min(fork_at_step) clamp per level. Previously they walked exactly one parent level, dropping every step that originated on the grandparent.
• Closer-to-leaf rows still shadow ancestor rows (existing owned-over-inherited dedup invariant preserved). Cycle defense via visited-set; depth cap of 64.
• Track 1 bump 0.15.1 → 0.15.2 across the 5 Rust-track files; Track 2 bump 0.17.0 → 0.17.1 (python/ was touched and 0.17.0 is published).

Repro (pre-fix)

On dev1 session ray-agent-7bea73fa (2026-05-25): a replay timeline forked from an edited-fork forked from main rendered as 1 step in the dashboard instead of all 4 inherited. Same shape on every 3+ level fork chain.

Test plan

• 4 new unit tests in rewind-replay cover 3-level inheritance, both clamp directions, and span symmetry — all fail before the fix, pass after.
• New web integration test test_steps_endpoint_walks_three_level_inheritance exercises the /api/sessions/{id}/steps endpoint through a 3-level fork chain.
• cargo test --workspace — all suites green (0 failed across 17+ test binaries).
• cargo clippy -p rewind-replay --all-targets -- -D warnings — clean. (rewind-web has pre-existing clippy errors in unrelated test files that were not introduced by this change.)
• Removed the stale "only walks one parent level up" caveat in test_fork_and_edit_step_inherited_step.

Version bumps

Per CLAUDE.md decision tree:

• Track 1 (workspace v0.15.1 released; crates/ touched): Cargo.toml, Cargo.lock, python/rewind_cli.py, python-mcp/rewind_mcp_cli.py all 0.15.1 → 0.15.2; python-mcp/pyproject.toml 0.13.11 → 0.13.12.
• Track 2 (rewind-agent 0.17.0 on PyPI; python/rewind_cli.py touched): python/pyproject.toml and python/rewind_agent/__init__.py 0.17.0 → 0.17.1.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
rewind	Ready	Preview, Comment	May 25, 2026 9:15am

[risjai]

Review notes after reading the full diff + running tests:

• ✅ Ran cargo test -p rewind-replay (20 passed).
• ✅ Ran cargo test -p rewind-web --test recording_api_tests test_steps_endpoint_walks_three_level_inheritance (passed).

One potential regression to consider:

MAX_ANCESTRY_DEPTH = 64 introduces a hard failure for any valid acyclic fork chain deeper than 64 levels. Since cycle detection is already handled via visited IDs, this cap may reject legitimate long-lived edit/fork histories and surface as endpoint failures for /steps or /spans.

Could we either (a) remove the hard cap, or (b) make it a soft guard + add a regression test for a >64-depth acyclic chain?

[risjai]

Good catch — applied option (a) in 8359f47.

The visited-set already rejects cycles, and while cursor.is_some() consumes one new id per iteration from a finite timelines.len() pool, so the walk is naturally bounded. The 64-level cap can only false-positive on legit deep histories, never catch a real cycle the visited-set would miss.

[risjai]

Re-checked after 8359f47.

No new blockers from my side:

• MAX_ANCESTRY_DEPTH cap removal addresses the only regression risk I called out.
• cargo test -p rewind-replay passes (20/20).
• cargo test -p rewind-web --test recording_api_tests test_steps_endpoint_walks_three_level_inheritance passes.

Remaining warning seen in rewind-web test compile (unused variable: root_tid around recording_api_tests.rs:1898) appears pre-existing/unrelated to this PR.