Development validation loop #

Status: operator-approved process direction; bounded opt-level 0/1 experiment completed successfully (2026-09-08). Current stable is the practical primary toolchain. Cargo defaults remain unchanged; opt1 is an explicit option for repeated expensive tests. Do not interrupt or change profiles during an active frozen checkpoint gate.

Default tiers #

1. Implementation feedback: check the affected crates and run the relevant named tests. Reuse a persistent stable target directory for the chosen profile rather than rebuilding dependencies in a new directory for every worker. Separate directories remain appropriate for deliberate compiler/profile comparisons or real isolation needs.
2. Impact integration: run affected cross-crate/process cases or targets. A unit test does not replace real process/durability evidence. Start with the relevant named process case when possible rather than rerunning every scenario.
3. Frozen checkpoint: after integration is coherent, run complete target coverage once on frozen source, plus required optimized build/replay and independent review. The existing metadata-derived partition auditor ensures no targets are omitted. Completed audited partitions are the full suite; do not then rerun an equivalent monolithic workspace command locally.
4. Review fixes: rerun affected checks. Repeat broader gates only when the change can affect them; retain completed evidence on unchanged source. Record exactly which source/version each result covers.

No speedup by ignored regressions, weakened assertions, disabled bounds/overflow/debug checks, fabricated state/history, global mutable test worlds or reduced persistence coverage. CI/platform checks are distinct evidence; do not misrepresent configuration as execution.

Observed baseline and likely cost centers #

The Wield CLI library checkpoint passed235 tests with1 existing ignored in1675.09s (27m55s), using the default unoptimized test profile. An earlier221-test run took1560.96s. A live process observation during the Wield run showed active CPU usage around five cores and about1.6GiB RSS, not an idle deadlock. Rust's over60-second notices are progress warnings, not individual test timings.

Read-only source audit found repeated full-world work in GameSession::from_identity/from_decoded, Story checkpoint proof and replay, canonical scans before Application::save, and process tests intentionally launching independent managed roots/binaries. These are plausible cost centers grounded in call paths, not a completed profiler attribution. Process isolation/durable milestone tests are deliberate coverage, not automatically a fixture-design defect.

Measured opt-level experiment (#57) #

On gameplay baseline 3eb764a, stable rustc/Cargo 1.98.1, x86_64 Linux, three existing tests ran twice per profile, serially with alternating profiles. All 12 exact libtest executions passed; no assertions or fixtures changed. Both DEV and TEST profiles explicitly set opt-level 0/1, debug assertions true and overflow checks true, retaining debuginfo 2. Separate initially absent target directories isolated builds. Direct libtest execution excluded Cargo overhead.

Build-only totals for CLI lib, phase11_gate and Save lib were 76.33s opt0 versus 184.65s opt1, an additional 108.32s. Later targets reused same-profile dependencies. These are empty-target builds, not OS-cold-cache measurements. Two samples per pair, fixed build order and uncontrolled cache effects limit generalization. Incremental rebuild costs, peak RSS and profiler attribution were not measured; do not extrapolate a full-suite time from these three tests.

Cargo JSON artifacts confirmed optimization level, enabled assertions/overflow checks and debug info for all test executables and the actual non-test hexwork child. Verbose compiler commands confirmed each phase11 harness's compile-time CARGO_BIN_EXE_hexwork selected its corresponding isolated binary; tests/support/mod.rs uses that value directly. The Save test re-encodes the same genuine v4 Classic/Story runs and replay byte-identically in both profiles. Independently allocated run IDs were not compared as world equivalence.

Parent checked the machine-readable ledger's six successful builds and twelve successful samples, recorded timings and compiler/artifact evidence. Local ignored evidence is under target/dev-loop-ab-57/ (RESULTS.md, report.json, runner and complete logs); these local files are not shipped repository artifacts.

Practical decision #

Keep default Cargo profiles unchanged until incremental-edit costs justify a global switch. For repeated expensive test work, opt1 is now a measured option using one persistent target directory. Apply both DEV and TEST settings so process children match, and retain the explicit safety checks. For example:

TMPDIR=/var/tmp CARGO_TARGET_DIR=target/dev-loop-ab-57-opt1 \
CARGO_PROFILE_DEV_OPT_LEVEL=1 CARGO_PROFILE_TEST_OPT_LEVEL=1 \
CARGO_PROFILE_DEV_DEBUG_ASSERTIONS=true CARGO_PROFILE_TEST_DEBUG_ASSERTIONS=true \
CARGO_PROFILE_DEV_OVERFLOW_CHECKS=true CARGO_PROFILE_TEST_OVERFLOW_CHECKS=true \
cargo +stable test --locked -p hexwork-cli --lib \
  session::tests::exact_run_restore_retains_state_history_and_continuation -- --exact

The target name reuses the measured local cache; new checkouts build it once. Do not alternate different flags in the same persistent directory or change a running checkpoint's profile. This option does not replace the focused/impact/frozen-checkpoint tiers above or justify another unchanged full-suite run.

No new general test framework is required. Immutable world-plan reuse may be investigated later only with correct seed/content keys and isolation; shared mutable runtime fixtures are not an acceptable shortcut. No additional legacy-toolchain project is part of this workflow work.