[fdaviddpt]

Closes #475

What was broken

scripts/rebuild-tsv.sh's ambiguous-vocabulary-keyword collision report built layerdir (derived from CLAUDE_PROJECT_DIR/VOCAB_BASE) and passed it to awk via -v layerdir="$layerdir". awk's -v decodes backslash escapes in the value it receives, so on a backslash-bearing CLAUDE_PROJECT_DIR (a Windows-shaped checkout path) the path arrived mangled inside the awk program, every bytesof() getline against it failed to open, and every entry in that vocabulary layer was reported "unmeasured" instead of its real byte count. Same defect class #424 fixed for JIT_BASE/tools_base/vocab_base and #437 fixed for GENERIC_WORDS_FILE/GENERIC_WORDS_FILES, in this same file -- #475 was filed because this one site was missed by both earlier fixes.

The fix

layerdir now travels into the awk program via ENVIRON["layerdir"] (exported as a plain bash variable in front of the awk invocation) instead of -v layerdir=..., matching the existing GENERIC_WORDS_FILES/ENVIRON pattern already used a few hundred lines below in the same file. The layer variable (a cosmetic label built by jit_report_name(), which only ever returns [A-Za-z0-9._-] or a withheld placeholder) was deliberately left on -v -- it can never carry a backslash and is not part of this defect class.

Test-first evidence

New case "475" added to tests/test-entry-bytes.sh, inside the existing per-awk-engine loop (awk/gawk/mawk). Builds a project root under a directory containing a literal backslash, writes two vocabulary entries sharing a keyword, forces the collision report, and asserts the collision still fires (positive control) while the stderr report does NOT contain "could not be measured" (negative control -- the literal text emitted on the bug).

RED: ran by hand against the pre-fix code (single git stash of scripts/rebuild-tsv.sh, popped immediately after) -- case 475 FAILED on awk and gawk, PASSED on mawk (expected engine variance; mawk's -v does not decode escapes the same way, matching this file's own documented #77/#78 variance shape).

GREEN: after restoring the fix, all three engines pass; full file reports PASS: 345 FAIL: 0.

Also ran: python3 .oss/assemble_changelog.py --check (ok, 3 fragments), tests/test-line-citations.sh (exit 0), shellcheck -x scripts/rebuild-tsv.sh (only a pre-existing, unrelated SC1091 info note, present regardless of this diff). Did not run bash tests/run-all.sh -- CI is the authority across its 3-platform matrix; ran the targeted suite plus two cross-cutting guard scripts (lane_setup.py), both of which reported not-in-this-repo/uncovered for this bash project.

Follow-up in this branch

trap.d/424.rebuild-tsv-v-escape.md was found stale during self-review -- still describing both its items (this one, and the already-fixed GENERIC_WORDS_FILE/wf site) as open. Removed in a second commit, since both are now resolved.

Adjacent, not fixed

A second -v-passed value exists at scripts/rebuild-tsv.sh's module-prefix-mapping report: -v file="$filename", used only in a printf into a different TSV, never opened via getline. Below the bar, not fixed: $filename is a basename() result, and a literal backslash cannot appear in a filename component on a real Windows filesystem (backslash is the Windows path separator, forbidden inside a filename), unlike layerdir, which is a full directory path where backslash legitimately appears as a separator -- so the defect class this issue fixes is not reachable at that site.

docs_targets

README.md checked for layerdir/bytesof/"Ambiguous vocabulary" -- no mention. Internal maintainer-tool reporting fix with no user-facing behavior change; no-change-needed.

[AI-generated]