v0.7.5

Fixed — the PowerShell -ArgumentList array form evaded the black gate (#124)

Start-Process powershell -ArgumentList '-ExecutionPolicy','Bypass','-WindowStyle','Hidden',...
rated green while the semantically identical space-separated spelling rated
black. Both obfuscation lookaheads required whitespace between a flag and
its value (\s+), but PowerShell's idiomatic array form separates them with a
quote-comma-quote. This is the deterministic hard-deny gate, not an advisory
flag, and it is load-bearing: @askalf/agent's shell-mode dispatch is firewalled
through the warden hook, so a miss here is unfirewalled execution on the host.
Proven live on a fleet device — the comma-array form completed, the
space-separated form was blocked, same box, same PID, 7 minutes apart.

• Two separator classes, split by left context — not one widened class. The
  obvious fix (\s+ -> ['",\s]+ everywhere) creates a false-positive class: a
  bare comma-space then bridges two ordinary prose words, so English like
  "net, localgroup and add are keywords" newly rates black. Instead FLAGSEP
  follows a dash-flag, where a bare comma is safe because prose never writes
  -windowstyle,hidden; ARGSEP sits between two bare words, where the comma
  must be quoted on both sides to count.
• Short-flag array spellings ('-ep','bypass','-w','hidden','-enc') are covered
  by the same arms. Pinned by test/argument-list-separator.test.mjs.

Fixed — -f matched the CLI flag, not just the format operator (#125, #123)

$t = "build -f Dockerfile -t app ."; & $docker $t smelled of obfuscation and
paid a needless judge call. OBFUSCATION_RE's -f\b arm keys on a {N}
placeholder now, which is not a heuristic: the format operator cannot assemble
without one ("a" -f $x evaluates to "a"), so a placeholder appears in every
real evasion and in no flag usage. Verdict-neutral — the smell routes to the
judge and never blocks — so this cuts wasted judge volume, not coverage.
#123 routes genuine PowerShell string-assembly obfuscation to the judge.

Corpus unchanged at 298 samples: deterministic recall 100%, precision
100%. 310 tests pass; ReDoS bench worst case 0.89ms against a 25ms threshold.

Note on distribution

npm remains a pointer stub (@askalf/redstamp 0.0.3) and NPM_PUBLISH_ENABLED
stays off — see 0.7.4. This release ships as the git tag plus the signed GitHub
release with Sigstore provenance, which is the only distribution channel. There
is no npm 0.7.x line and this release does not create one.