walk-g1-g7 — build the tower with only form-cli

The discipline: only form-cli subcommands. No hand-written recipes, no Edit/Write,
no remote LLM unless the loop itself escalates. Local oracle is ollama run coder
(or qwen2.5:72b); the remote tier is claude -p and every use of it is COUNTED.

The loop, per open step

form-cli roadmap                     # the plan; read the next open step + its spec
form-cli asm "<C for the construct>" # learn LLVM's lowering offline (clang, local)
form-cli close "<id>" "<spec informed by the asm>" "<assert>" "<expected>" "ollama run coder"
# close: a LOCAL oracle drafts the recipe; the kernel validates it against the assert.
# if it validates -> register a band + run: cd form/form && ./validate.sh ... (four-way) -> ship.
# if it does not -> record WHY, do not hand-write the fix; the point is to find what
#   form-cli itself still lacks.

What to measure and report

• remote calls: count every time the loop used claude -p (target: 0).
• shipped: how many steps produced a four-way-validated recipe via form-cli alone.
• for each not shipped: the exact reason (the kernel error, the missing capability),
  and the smallest fix to form-cli that would let it ship next time.

The honest frame

A step "shipped via form-cli" only if a form-cli command (not the operator) produced
the validated recipe. If the operator had to write the recipe by hand, that step did
NOT ship via form-cli — record it as a gap in form-cli and name the fix. The report's
value is the gap list + fixes, not a green count.