perf(sdk): stop cloning turn streams in analyze hot paths

[willwashburn]

Summary

Closes #325. Aggregate per-session / per-file groups in the analyze hot paths by &TurnRecord instead of cloning the full record. TurnRecord carries a Vec<ToolCall> plus optional file lists, so these clones are not cheap.

• overhead.rs: per-file filter now collects Vec<&TurnRecord> and routes through a new internal attribute_claude_md_refs. The public attribute_claude_md is unchanged — it adapts via a single ref collect. Biggest win: every Claude turn used to be cloned ~2× and every Codex/OpenCode turn ~1× across the default file candidates.
• hotspots.rs: the per-session group is IndexMap<String, Vec<&TurnRecord>>. Private attribute_session and index_tool_results now take &[&TurnRecord].
• quality.rs: per-session aggregation borrows turns; private infer_outcome_refs / compute_one_shot_rate_refs do the work. The public infer_outcome and compute_one_shot_rate keep their existing signatures and adapt via a single ref collect.
• compare.rs: hoist the model-name clone above the three entry/or_default calls so we allocate one owned String per new model (instead of four per turn). by_model_category is now the canonical insertion gate since model_set and model_totals may be pre-seeded by the model filter.

The describe_applies_to follow-up the issue mentioned is already in main (uses SourceKind::wire_str returning &'static str).

Test plan

• cargo build --workspace
• cargo test --workspace — all 657 SDK tests pass, plus CLI / sdk-node suites
• cargo clippy --workspace --all-targets — no new warnings from this change

🤖 Generated with Claude Code

---

Generated by Claude Code

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 4e8454cc-a435-49d2-b535-58016f12e430

📥 Commits

Reviewing files that changed from the base of the PR and between 67d7255 and 23a4003.

📒 Files selected for processing (6)

• CHANGELOG.md
• crates/relayburn-sdk/src/analyze/claude_md.rs
• crates/relayburn-sdk/src/analyze/compare.rs
• crates/relayburn-sdk/src/analyze/hotspots.rs
• crates/relayburn-sdk/src/analyze/overhead.rs
• crates/relayburn-sdk/src/analyze/quality.rs

---

📝 Walkthrough

Walkthrough

This PR eliminates TurnRecord cloning across the analyze module's memory-intensive hot paths by switching per-session and per-file aggregations to use borrowed references. The refactoring introduces a reference-based Claude Markdown attribution helper and updates hotspots, quality, compare, and overhead modules to operate on Vec<&TurnRecord> views, reducing working-set memory pressure while preserving all behavior.

Changes

TurnRecord reference-based aggregation optimization

Layer / File(s)	Summary
Claude Markdown reference-based attribution refactoring crates/relayburn-sdk/src/analyze/claude_md.rs	Introduces internal attribute_claude_md_refs(files: &[ParsedClaudeMd], turns: &[&TurnRecord], pricing: &PricingTable) helper to accept borrowed turn slices and build per-session indices without cloning. Public attribute_claude_md pre-borrows input turns and delegates to the new helper.
Hotspots per-session indexing optimization crates/relayburn-sdk/src/analyze/hotspots.rs	Per-session indexing now stores Vec<&TurnRecord> references instead of cloning. Updates attribute_session and index_tool_results helper signatures to operate on borrowed turn slices (&[&TurnRecord]).
Quality analysis reference-based aggregation crates/relayburn-sdk/src/analyze/quality.rs	Groups turns per session as &TurnRecord reference slices and delegates to new internal infer_outcome_refs and compute_one_shot_rate_refs helpers. Public wrappers build borrowed views internally; public signatures unchanged. Updates ending_role and trailing_failure_streak helpers to operate on reference slices.
Compare table accumulation optimization crates/relayburn-sdk/src/analyze/compare.rs	Main aggregation loop now derives model as &str and applies allow-list checks on the borrowed value. Population of model_set/model_totals/category_set occurs only on first insertion into the canonical map; accumulation uses get_mut on pre-created entries instead of entry().or_default() with per-iteration cloning.
Overhead attribution integration crates/relayburn-sdk/src/analyze/overhead.rs	Per-file filtering now collects matching turns as Vec<&TurnRecord> and calls attribute_claude_md_refs instead of the prior value-based attribute_claude_md with full clones. Eliminates the 3× overhead turn cloning bottleneck.
Changelog update CHANGELOG.md	Documents hot-path analysis optimization: per-session/per-file grouping now aggregates by reference instead of cloning TurnRecords, reducing working-set memory for overhead, hotspots, quality, and compare while preserving behavior.

🎯 3 (Moderate) | ⏱️ ~25 minutes

Possibly related PRs

• AgentWorkforce/burn#291: Modifies hotspots attribution implementation and changes how turns are handled in aggregation.
• AgentWorkforce/burn#398: Refactors hotspots functionality including attribute_session and aggregation rollups.
• AgentWorkforce/burn#349: Modifies compare table construction path and build_compare_table integration.

🐰 A hop through memory springs so light,
No clones of turns from morn to night,
By ref we bound, the workset freed,
Hot paths now run with newfound speed!
Each turn's a view, no dupes remain—
Less alloc joy, less malloc pain. 🌿

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title 'perf(sdk): stop cloning turn streams in analyze hot paths' is concise, specific, and clearly summarizes the main performance optimization in the changeset.
Description check	✅ Passed	The description is comprehensive and directly related to the changeset, explaining the rationale, implementation details across all modified files, test results, and linking to issue #325.
Linked Issues check	✅ Passed	The PR fully addresses all coding objectives from issue #325: eliminates cloning in overhead.rs, hotspots.rs, quality.rs, and compare.rs by using &TurnRecord references; maintains behavioral parity; and passes all tests.
Out of Scope Changes check	✅ Passed	All changes in the PR are scoped to the objectives of issue #325: refactoring hot-path analysis functions to use references instead of clones. CHANGELOG.md addition documents the change appropriately.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch claude/submit-pr-review-b1sse

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 4 additional findings.