Beagle SCM

CFOLD-002: a tokenizer failure masks as "buffer full" and turns a mergeable file into a FALSE CONFLICT

DIFF-015 fixed the diff view: a fold failure now falls back to the plain-text lexer instead of erasing the file. The SAME mask survives on the merge side. jab/cfold.hpp:180 collapses every non-CFOLDBIG failure into cfold.fold: failed (buffer full?), and four JS callers gate on the substring "full", so a tokenizer derail is silently taken for an over-cap skip. Measured 2026-08-04: weave3(base, ours, theirs, "js") on verbs/mark/render.js bytes THROWS, the same call with ext "" merges (26269 bytes) — so jab get over an uncommitted edit to any slash-in-class .js file (DOG-021) reports a conflict that does not exist.

Input

Context

• Live throw site is jab/cfold.hpp:179-180, not jab/weave.hpp:189 — the latter is dead code that DIFF-015 named; only CFOLDBIG is forwarded distinctly.
• Gates matching the catch-all substring: verbs/get/get.js:1127, shared/checkout.js:272, verbs/patch/patch.js:219, shared/expected.js:137 (all verified present at #014aa381).
• Harm per gate: get and checkout set mg = null (bogus con row), patch emits failed, expected drops the EXPECTED provenance — all with no signal, exactly the DIFF-015 class.
• CFOLDBIG's own message ("the file is too big to weave (over 16 MB of history)") contains no "full", so every one of those gates RETHROWS where it meant to skip — latent, no be/ file reaches it today.
• Clean, do not touch: views/diff/diff.js:167,179,306 and core/loop.js:88 (grow-and-replay) gate correctly for their own purpose.
• Trigger is DOG-021: JST ends a regex literal at a / inside a char class, so any file with such a class folds JSTBAD.

Goals

• A fold/merge failure NEVER silently degrades a merge: either the file merges, or the operator is told in plain words what happened (README norms).
• jab/cfold.hpp throws carry the real ok64 code, so over-cap, over-16MB-history and tokenizer-fail are three distinguishable outcomes, not one string.
• Callers gate on the code, never on a substring of a catch-all message; the CFOLDBIG rethrow-instead-of-skip gap closes with them.

Constraints

• Touches two repos: jab/cfold.hpp (the throw) and ///be JS gates; jab/ is read-only for agents, so the C++ side needs gritzko or an explicit order.
• No parallel differ/merger: the ruled remedy shape is a retry-fold under libdog's TXTT (ext ""), as DIFF-015 landed for diff — a hand-rolled Myers differ was rm -rf-ed once for this.
• Repro-first: the false conflict must be watched failing in a be wt before any gate is opened up.

WIP

Design decisions

• OPEN QUESTION for gritzko: may MERGE fall back to TXTT like diff does? That trades token granularity for line granularity on the failing file — the 2026-07-29 ruling covered the diff VIEW only, and a merge ruling was not given.
• Until that is ruled, the safe half is the diagnosis: forward the code and make each gate say which of the three things happened.

TODOs

• Repro: jab get over an uncommitted edit to a slash-in-class .js file yields a bogus con row, watched failing first.
• jab/cfold.hpp: fold/merge throws carry the ok64 code; over-cap, CFOLDBIG and tokenizer-fail are distinct.
• Gates match the code, not "full": get.js:1127, checkout.js:272, patch.js:219, expected.js:137.
• CFOLDBIG reaches those gates as a rethrow today — cover it once the codes are forwarded.
• Drop or align the dead jab/weave.hpp:189 throw so it stops advertising the old message.
• After the ruling: merge-side TXTT retry, or an explicit refusal with a plain-words notice.

Blockers and bummers

• DOG-021 stays open meanwhile: any slash-in-class .js file folds JSTBAD, so the false conflict is reproducible on demand.
• The merge-granularity ruling above blocks the last TODO only; the forwarding work is unblocked.

Outcome

• (pending)