dss-state

name

dss-state

description

Measure the current state of the DSS Code Prime compiler. Runs a 104-probe C-feature battery through the real dss-code-prime CLI and reports every % axis — empirical C coverage, plan-23 full-C arc %, SQLite-readiness %, test suite, deferred-anchor closure, cross-target emit/run matrix — plus a velocity-based ETA for "compiles SQLite" and the full C23 arc. Use when asked how far along the compiler is, for a progress/state/status report, % complete, an ETA for full C or SQLite, or to re-measure after a /dss-cycle.

Computes a live state-of-the-compiler report by driving the real CLI: every probe in probes.mjs is compiled by dss-code-prime --compile for the host target and the produced binary is executed — so the percentages measure the whole pipeline (parse → semantic → IR → codegen → link → runtime), not the grammar. Static axes (plan-23 phase board, anchor registry, ctest log, git) are parsed from the repo.

All paths below are relative to the repo root.

Prerequisites

• Node ≥ 18 (verified with v22.12) and git on PATH — nothing else; the driver has zero deps.
• A built CLI. The driver auto-picks the newest of build/bin/dss/{Debug,Release}/, build-rel/bin/dss/, build-dbg/bin/dss/. To (re)build the MSVC Debug one:

cmake --build build --config Debug --target dss-code-prime

• Optional (Windows): WSL gives the ELF runtime witnesses. This box has qemu-aarch64 8.2.2 and the /usr/aarch64-linux-gnu sysroot inside WSL, which the arm64 leg needs (see Gotchas).

Run (agent path)

node .claude/skills/dss-state/driver.mjs

Full report to stdout in ~20–60 s (battery itself is ~2–7 s; WSL legs dominate). Variants:

node .claude/skills/dss-state/driver.mjs --quick                      # skip WSL → ~10 s
node .claude/skills/dss-state/driver.mjs --ctest                      # live ctest run instead of the cached log (+ ~2 min)
node .claude/skills/dss-state/driver.mjs --json out.json --md out.md  # also write machine/markdown copies

flag	effect
--quick / --no-wsl	skip WSL/qemu runtime legs (matrix shows emit-only)
--ctest	run the full ctest suite live (default: parse the cached build/Testing/Temporary/LastTest.log, timestamped in the report)
--json <p> / --md <p>	write the structured dump / the markdown report
--cli <exe>	use a specific compiler binary instead of newest-found
--jobs N	probe concurrency (default min(8, cores))
--keep	keep the temp probe dirs (%TEMP%/dss-state-<pid>/) for inspection
--strict	exit 1 if any miscompile/crash/toolfail was detected

Reading the report

• 🔴 MISCOMPILES is the highest-priority section: rc=0 compile but WRONG runtime result — the silent-miscompile class this repo treats as critical. (The battery caught a real one on its first run: ptr_swap_through — see Gotchas.)
• Funnel: each rejected probe is classified by the first letter of its diagnostic codes (P_* parse, S_* semantic, H/I_* IR, L/A/E_* backend) — so you see where each feature dies, not just that it fails.
• ETA model: remaining-cycle estimates per FC phase live in PHASE_CYCLE_EST (driver.mjs, calibrated on FC1–FC4 actuals); velocity = completed-phase cycles / calendar days since PLAN_ARC_START, with a SUSTAIN_FACTOR haircut for the steady scenario. SQLite milestone = cheapest of two routes (Win64-PE vs SysV-ELF prerequisite phase sets). All three scenario dates + the math are printed; treat them as extrapolations.
• If plan-23 gains/renames FC phases the driver fails loud (PLAN DRIFT row / fatal) — update PHASE_CYCLE_EST and the SQLITE_PATHS sets to match.

Extending the battery

Add probes to probes.mjs — { id, cat, expect, src } (+ optional stdout, multi-TU files). Style rules are at the top of that file; the two that matter: declare-then-assign (init-at-decl has its own probe), and never use Bool-arithmetic ((a<b)*20 hits a known semantic gap) — accumulate via if. New probes need no driver change.

Test

node .claude/skills/dss-state/driver.mjs --quick

Expect: a battery of 104 funnel line, the category table, and exit 0. (--strict also exits 0 since 2026-06-12 — the ptr_swap_through miscompile is FIXED, see below.)

Gotchas

• Artifacts are named after the FIRST source file's stem (--output dir + main.exe/main). Every probe writes its source as main.c for this reason; a canonical.c once produced canonical.exe and the matrix read 0/5 emit. Keep battery sources named main.c.
• The exe is a thin main over dss-code-prime.dll — rebuilds usually relink only the DLL, so the staleness check compares the newest of exe+DLL against on-disk src/**/*.{cpp,hpp,…} (commit timestamps skew by minutes; .json configs are runtime-loaded and never stale it).
• DSS ELF executables are dynamic (every program imports libc exit), so bare qemu-aarch64 fails with Could not open '/lib/ld-linux-aarch64.so.1'. The driver passes -L /usr/aarch64-linux-gnu; without qemu or that sysroot in WSL the leg reports skipped (wsl-qemu runner unavailable) — install qemu-user + an aarch64 sysroot (gcc-aarch64-linux-gnu provides it) to light it up.
• Per-format features flip green per HOST (TLS arc example). The battery compiles probes for the HOST-NATIVE target — pe64 on Windows, elf64 on Linux. A feature landed per-format (thread_local: ELF x86_64 landed TLS C1, arm64-ELF landed C2; PE lands C3; Mach-O C4) therefore shows c11_thread_local GREEN on a Linux host but 0x8015-RED on the Windows host until its C3 leg lands — the red is CORRECT (fail-loud on the un-landed leg), not a regression. Judge such probes on the leg that owns them.
• Git velocity is squash-blind: PRs squash to one main commit, hiding per-cycle commits. The plan board is the primary velocity source; the ^v0.0.2 commit count is shown only as a floor.
• Multiple build dirs — the ctest axis follows the CLI's dir, not build/. This box has both an MSVC build/ (often STALE/broken — semantic_analyzer.cpp overflows COFF sections without /bigobj) and a healthy Ninja build-dbg/. The driver auto-picks the newest CLI (usually build-dbg/) and now derives the ctest dir + LastTest.log from that same path (buildInfoOf()), so a single-config Ninja dir gets no -C. Before this fix the test axis hardcoded build/ and reported garbage (a 2/2 partial log; a 145/448 from the broken MSVC dir) while the real suite was 448/448 in build-dbg. To rebuild the active one: cmake --build build-dbg --target dss-code-prime (Ninja, single-config — no --config).
• The ptr_swap_through miscompile is FIXED (2026-06-12): the probe's exit-34 was NOT a pointer bug — the swap compiled perfectly; return x - y + 4; parsed RIGHT-associative (x - (y + 4)) because same-precedence infix chains nested rightward in the Pratt walker (D-PARSE-INFIX-ASSOCIATIVITY-STRUCTURAL ✅, registry). The probe now passes, MISCOMPILES is empty, and --strict exits 0. Lesson for probe triage: a miscompile probe's NAME frames a hypothesis (pointers) — the disassembly, not the name, localizes the defect.
• PowerShell quirk while probing by hand: piping the CLI through Select-Object -First N can kill it before it exits — $LASTEXITCODE comes back empty. The driver is unaffected (Node captures properly); just don't trust rc= after a truncated pipe.

Troubleshooting

• fatal: no dss-code-prime binary found — build one: cmake --build build --config Debug --target dss-code-prime (MSVC multi-config: the --config Debug matters; plain cmake --build build builds no config).
• fatal: plan-23 parse found only N FC rows — plan-23's §0.1 table format changed; fix parsePlan23() + recalibrate PHASE_CYCLE_EST.
• Matrix row WRONG exit … — that's a real cross-target regression, not a harness issue; the same canonical program just ran correctly elsewhere.
• Test-suite axis says no data — no build/Testing/Temporary/LastTest.log yet; run with --ctest once.