Cover the argument parses that produce a different run rather than an error

[hahwul]

91 examples over six seams that had none, picked by what they cost when wrong. Every one of them can be wrong and still exit 0: a flag that mis-parses does not raise, it sends different bytes, stores the rule somewhere else, or no-ops a mutation while reporting success.

file	examples	what it pins
spec/cli/run/rewriter_spec.cr	23	gori run rewriter — 710 lines that had no spec at all
spec/cli/run/fuzz_args_spec.cr	22	the payload flag parsers fuzz and discover share
spec/settings_scan_rules_spec.cr	20	the global scan-rule library
spec/cli/run/probe_helpers_spec.cr	14	probe's flag parses + rule-id decoding
spec/cli/run/list_leftovers_spec.cr	8	the discarded-verb refusal twelve list dispatchers share
spec/rules_spec.cr	+4	Rules.normalize_shape

How these were picked

Crystal has no working coverage tooling (kcov is Linux-only, crystal-coverage is dead), so this used an identifier-index heuristic instead: tokenise all of spec/, then count how many of each source file's def names appear in that set.

The useful result was the negative one. gori's public surface is already essentially covered — of 581 source files, exactly one (settings/scan_rules.cr) had its entire public def self. surface unmentioned anywhere under spec/. So the remaining gaps are in two places, and all six seams here come from them:

• the private def self. helpers inside cli/run/*.cr, whose names never appear in a spec because the parsing and formatting are all private
• functions whose own comment says they were split out to be spec-able — Run.list_leftover_error says exactly that and had zero coverage

What each one costs when wrong

rewriter — the scope letter leads a rule row because the project and global stores number independently, so #3 alone does not say which rule the next command addresses. The JSON projection is what a script reads, and default_enabled beside enabled is the only way that script can tell an override from a library default. Also the two refusals add makes before a rule reaches a store: an unparseable stub (it would answer every matching request with gori's own 502) and --part=ws on an op that is not replace (normalising it would move the rule to a different PROTOCOL, silently).

fuzz/discover flags — each has already been wrong in a way that produced a quietly different payload set: --numbers -10--5 split on the first hyphen and lost its lower bound, --regex-replace /foo//bar/ dropped everything past the second delimiter, --brute split its charset on the wrong colon. Checked through what each source actually generates, since none of them expose their fields.

list_leftover_error — what stops gori run rewriter --project=t1 rm 1 from listing the rules and exiting 0 with the delete discarded, and what keeps a lone list from being called an unknown subcommand by a message that lists list among the verbs.

probe rule ids — custom_p_12 is a row in this project's DB, custom_g_12 lives in settings.json. Reading the global one as a row id would send a delete at whatever project row happened to share the number.

scan_rules — its parse is the boundary a hand-edited settings.json crosses into the match engine, which reads side/region/kind/severity as though they were enums, so the clamp is the contract. A rule missing id/title/pattern is dropped rather than defaulted, because an invented pattern would scan live traffic for it. The delete of the LAST rule is checked on disk, not through a reload: an emptied section stops being serialized, which is the exact shape that has made save's merge copy a stale block forward before.

Limits

The abort branches remain out of reach — abort calls exit, which would take the suite down — so these cover success paths only, the limit spec/cli/run/links_spec.cr already works under. Reaching the refusal paths would need a harness that runs the built binary as a subprocess; there isn't one today, and adding it was left out of scope.

Verification

• crystal spec: 8592 examples, 8 failures — the same 8 as before this branch (project_registry ×5, capture_status ×3), all from a gori already listening on :8070. No new failures.
• crystal tool format --check and ameba: clean on every file touched.
• Source is untouched; this is spec-only, plus a refresh of the now-stale split-status header in spec/cli_run_spec.cr.