fix: shard subprocesses inherit CWD and profile for correct backend selection

[rysweet]

Summary

This PR contains two independent fixes to skwaq gym eval correctness.

---

Fix 1: Shard subprocesses inherit CWD and profile for correct backend selection

Root Cause

During skwaq gym eval, shard subprocesses were selecting the Copilot backend instead of Azure AI Foundry. Traced to two bugs in spawn_shard() in crates/cli/src/commands/gym_cmd.rs.

Config loading code path

Config::load() → find_config_file() searches:

1. skwaq.toml in current working directory
2. .skwaq/config.toml in CWD
3. ~/.skwaq/config.toml

If no file is found → Config::default() → default_llm_backend() returns "copilot".

Bug 1: Missing .current_dir(skwaq_root)

spawn_shard did not call .current_dir(), so shard subprocesses inherited the parent's CWD. If skwaq gym eval was invoked from a directory other than the repo root, Config::load() in each shard would not find skwaq.toml and fall back to Config::default() — selecting Copilot regardless of what the config file specifies.

Fix: Set .current_dir(skwaq_root) so shards always load config from the repo root.

Bug 2: --profile not forwarded to shards

When running gym eval --profile azure, the profile's LLM backend override (reasoning = "azure") was applied in the parent process but not forwarded to shard subprocesses. Each shard called Config::load() directly from skwaq.toml without the profile overlay, potentially selecting a different backend.

Fix: Forward --profile <name> to shard subprocess args when a profile is active.

Changes (Fix 1)

• crates/cli/src/commands/gym_cmd.rs: spawn_shard gains two new parameters (skwaq_root, profile); both call sites updated (initial spawn + retry-on-early-death).

---

Fix 2: CyberGym adapter — prevent source_tree_root() over-descent; WARN on patch fallback

Root Cause

CyberGym cases are often packaged as case_dir/<project-1.2.3>/src/vuln.c. The helper source_tree_root() was supposed to strip the packaging wrapper (project-1.2.3/) to find the project root. However, single_child_dir_without_sources() would descend one more level into src/ because it had no guard against conventional source-layout directories. This caused patch-relative paths like src/parser.c to fail resolution, producing false negatives.

Additionally, when both patch_affected_files and paired_case_affected_files failed, the shallow fallback scan ran silently with no log output, making false negatives invisible.

Changes (Fix 2)

• crates/gym/src/adapters/cybergym.rs:
  • Added SOURCE_LAYOUT_DIRS constant listing conventional source directories (src, lib, include, source, Sources, core, main, libs, includes).
  • single_child_dir_without_sources() now returns None when the sole child directory name is in SOURCE_LAYOUT_DIRS, keeping source_tree_root() at the project root instead of descending into the source tree.
  • Added three regression tests:
    • test_source_tree_root_stops_at_project_root_not_src
    • test_source_tree_root_stops_at_conventional_lib
    • test_patch_affected_files_resolves_src_relative_paths

---

Verification

Fix 1

• cargo build -p skwaq ✅
• cargo clippy -p skwaq --tests -- -D warnings ✅

Fix 2

• cargo test -q -p skwaq-gym ✅ (all 85 tests pass including 3 new regression tests)
• cargo clippy -q -p skwaq-gym --tests -- -D warnings ✅ (zero warnings)

CI

CI is triggered on push. Any CI failures will be documented here as root causes are identified.

[rysweet]

Superseded by #468. This older branch/PR path picked up unrelated scope and workflow residue; the focused replacement PR carries the backend-selection + CyberGym fixes with QA/docs evidence.