CUE_BYPASS=1 does not do what both docs say it does

[NagyVikt]

CUE_BYPASS=1 is documented in two places as a full escape hatch:

• docs/shell-install.md:77-82 — "Skip cue entirely … makes cue exec the real binary directly without touching the profile, materializer, or config dir. Use it when you need a raw claude session for debugging."
• docs/launch.md:207 — "CUE_BYPASS=1 claude — exec the real binary directly; no resolve, no materialize, no profile."

Nothing implements that. Every reader in src/:

Location	What it actually does
launch-loader.ts:160	returns null — suppresses the startup spinner
launch.ts:142 (added in #133)	skips the CUE_SMART_SUBSET argv fold
claude-classifier.ts:168, discover.ts:1522, discover.ts:2240, profile-draft-skill.ts:107	set it on a spawn; none read it

So CUE_BYPASS=1 claude --version still runs the full pipeline: resolve, materialize, profile, config dir. The documented debugging escape hatch does not exist.

Why it matters

This is what made #133 possible. claude-classifier.ts set CUE_BYPASS=1 on its spawn and its docstring stated the flag made cue's shim "transparent" — a reasonable belief given the docs. It wasn't, so the spawn re-entered cue launch and folded its own argv into the next classification prompt. Measured on a live machine: a 67KB command line with the prompt template repeated 10 times, and ~2.9GB across 7 concurrent classifier processes.

#133 fixed the loop but deliberately did not widen CUE_BYPASS — that is a behavioral change that didn't belong in a memory fix. It added a second narrow reader instead, which arguably makes the flag's meaning muddier, not clearer.

Options

1. Implement the contract. Short-circuit at the top of launch.run(): when CUE_BYPASS === "1", resolve via findRealAgentBin() and exec with the passthrough args, returning its exit code. Matches both docs, and makes the launch.ts:142 guard redundant (keep it as defense-in-depth or drop it).
   • Check first that discover.ts / profile-draft-skill.ts want a raw agent — they appear to (they spawn claude -p for classification/drafting), but confirm.
2. Correct the docs to describe the three narrow behaviors the flag actually has, and rename or split the flag.

(1) is preferable — the documented behavior is the useful one, and it removes a whole cue launch boot from every classifier spawn.

Note

MAX_LAUNCH_DEPTH = 3 bounds the process nesting, so this is a waste/correctness issue rather than a runaway-recursion one.