perf: cut Windows CI test fixture overhead

[ScriptedAlchemy]

Summary

Windows CI junit data from the latest green master run (28503571229) shows 10 shards at 11-13.5 min each with 3182 tests summing to ~235 CPU-minutes, and a hard ~15s fixture floor: any test that initializes a TraceDecay project or DB costs 15-65s on Windows (p50 for mcp_handler_test is 22.4s; schema-only tests run in 0.02s). This PR attacks the floor from four directions:

• Optimized SQLite in test builds (Cargo.toml): the bundled SQLite C sources in libsql-ffi were compiled at -O0 in dev/test builds. Scoped opt-level = 2 overrides for libsql-ffi/libsql-sys/libsql-rusqlite cut measured per-test time 2.8x on Linux (1.69s to 0.60s); the MSVC delta should be at least as large.
• TRACEDECAY_SQLITE_UNSAFE_FAST=1 escape hatch (src/db/connection.rs): tests/CI-only env var that switches every connection to journal_mode=MEMORY + synchronous=OFF, eliminating the per-transaction rollback-journal create/write/fsync/delete cost of the Windows DELETE+FULL pairing. In-memory journals never enter WAL mode, so the Windows WAL-teardown crash stays avoided. Tests that assert durable journal/synchronous modes pin the var off via env guards.
• Runner tuning (ci.yml, windows-test-shard only): disable Defender real-time monitoring plus exclusion-path fallback for workspace/temp/cargo dirs, redirect TMP/TEMP off the slow C: user temp onto the RUNNER_TEMP drive, and export the unsafe-fast env for the nextest step.
• Fixture trims in the slowest junit offenders: drop dead session-store seeding from scheduler gate tests (gate is evaluated before any store is opened); pin TRACEDECAY_GLOBAL_DB at each skill-writer runner test's own store (evidence building was opening the shared per-user global DB, contended by every parallel test process — the 9-of-10-slow signature of that binary); pre-create dashboard GlobalDb-schema stores from a cached empty template so server startup and seeding skip full schema creation (the curation restart test starts the server twice); seed branch-drift tracking metadata directly instead of building and syncing a branch DB only to delete it.

Verification

• cargo fmt --all -- --check, git diff --check, cargo clippy --workspace --all-targets --locked -- -D warnings
• cargo nextest run --workspace --profile ci --locked with TRACEDECAY_SQLITE_UNSAFE_FAST=1: 3272 passed, 0 failed
• Mode-asserting suites green both with the var set and unset (migration_test, profile_storage_migration_test, corruption_test, src/db/connection.rs unit tests)
• scripts/check-conventional-commits.sh origin/master..HEAD
• One earlier full-run failure (dashboard_session_and_skill_runs_emit_activity_when_evidence_is_unavailable) reproduces identically with these changes stashed under heavy machine load and passes reliably otherwise — pre-existing load-sensitivity flake, not a regression.

Test plan

• Windows shards green on this PR's CI run
• Compare shard wall-clock times and junit per-test times against master run 28503571229 (baseline: 11-13.5 min/shard, mcp_handler_test p50 22.4s)

[changeset-bot]

⚠️ No Changeset found

Latest commit: e8a0406

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR