Authorize frozen letter-mix on the public CLI as 0.4.0

.github/workflows/ci.yml

@@ -49,16 +49,45 @@ jobs:
           SH
           chmod +x "$test_bin/wl-copy"
           export PATH="$test_bin:$PATH"
+          expected_long="$(printf 'I do not agree and I cannot stay.\n' | python -c "from fuckmark.cycle8.letter_mix import apply_letter_alternating_mix; import sys; sys.stdout.write(apply_letter_alternating_mix(sys.stdin.read()))")"
+          expected_short="$(printf 'I do not agree.\n' | python -c "from fuckmark.cycle8.letter_mix import apply_letter_alternating_mix; import sys; sys.stdout.write(apply_letter_alternating_mix(sys.stdin.read()))")"
           for command in FuckMark Fuckmark fuckmark; do
             command -v "$command"
             capture="$RUNNER_TEMP/${command}.clipboard"
             export FUCKMARK_CLIPBOARD_CAPTURE="$capture"
             output="$(printf 'I do not agree and I cannot stay.\n' | "$command" --stdin --copy)"
             printf '%s\n' "$output"
-            test "$output" = "I do not agree and I cannot stay."
-            test "$(cat "$capture")" = "I do not agree and I cannot stay."
+            test "$output" = "$expected_long"
+            test "$(cat "$capture")" = "$expected_long"
           done
-          test "$(printf 'I do not agree.\n' | FuckMark)" = "I do not agree."
+          test "$(printf 'I do not agree.\n' | FuckMark)" = "$expected_short"
+          test "$(printf 'I do not agree.\n' | FuckMark --stdin --visible)" = "I do not agree."
+          expected_arg="$(python -c "from fuckmark.cycle8.letter_mix import apply_letter_alternating_mix; import sys; sys.stdout.write(apply_letter_alternating_mix('I do not agree.'))")"
+          test "$(FuckMark 'I do not agree.')" = "$expected_arg"
+          FuckMark --help >/dev/null
+          python - <<'PY'
+          from io import StringIO
+          from fuckmark.cli import main
+          from fuckmark.cycle8.letter_mix import apply_letter_alternating_mix
+
+          class Tty(StringIO):
+              def isatty(self):
+                  return True
+
+          copied = []
+          out = StringIO()
+          err = StringIO()
+          src = Tty("I do not agree.\n\nSecond paragraph.\n:done\n")
+          assert main(src, out, copied.append, error_stream=err) == 0
+          expected = apply_letter_alternating_mix("I do not agree.\n\nSecond paragraph.")
+          assert copied == [expected], copied
+          assert out.getvalue() == ""
+          ui = err.getvalue()
+          assert "Processing..." in ui
+          assert "Copied to clipboard" in ui
+          assert expected not in ui
+          print("interactive e2e ok")
+          PY
 
   calibration-readiness:
     name: Frozen calibration readiness

CHANGELOG.md

@@ -1,92 +1,22 @@
 # Changelog
 
-## v0.3.0 — Visible invariance, install and release hardening
+## v0.4.0 — Product CLI
 
-- Advanced the project version to 0.3.0 so the identity CLI is not published under the v0.2.0 contraction tag.
-- Restored Priority Zero: product output must keep the exact user-visible projection of the input.
-- Public CLI `release-cli-v4` fail-closes instead of applying contractions or other visible edits.
-- `release_transform_registry()` is the empty product-visible-invariance registry. Historical contraction/lexical/syntax/spacing catalogs remain replayable under explicit historical names.
-- Added `specs/fuckmark-user-visible-invariance-v1.contract.json` and Cycle 8 invisible-carrier research. No product-authorized carrier is promoted yet.
-- Reclassified Cycle 6 formal `NONZERO_RESIDUAL` 7/192 and Cycle 7 visible-edit stages as scientific evidence that is product-disqualified.
-- GitHub CI installed-CLI end-to-end now expects unchanged visible text, matching `release-cli-v4`.
-- Cycle 8 fixture compare reports visible-projection pass `20/20` for space-carrier arms.
-- Chromium `pre` screenshots: U+034F / U+FE00 pixel-equal; U+200C PNG bytes differ.
-- Detector-blind GPT-2 / SynthID development scoring on seeds `890000`, `900000`, and `910000` reduced watermarked detections for U+034F space-carrier x1 without unwatermarked inflation on these tiny 4-pair corpora. This is not confirmation and is not a product promotion.
-- Product contract loading uses the embedded v1 payload in installed wheels. Writers that replace pinned `specs/cycle8` tokenizer JSON refuse to run without a GPT-2 encoder; the detector harness still scores when tiktoken is absent. Chromium HTML evidence encodes less-than and greater-than as JSON unicode escapes so an ASCII script terminator cannot close the inline script.
-- Official install docs now use the tagged GitHub Release wheel plus `SHA256SUMS.txt`. In-repo installers verify that checksum, do not start the CLI, and do not use sudo. `d.q1z.org/mark` is no longer documented as the install path.
-- Release Engineering no longer auto-tags, auto-publishes, or deletes merged branches on `main` push. Publishing is `workflow_dispatch` only against an existing tag.
-- GitHub Actions `workflow_dispatch` values are passed through `env:` instead of being interpolated into `run:` scripts.
-- Interactive CLI copies to the clipboard only with `--copy`.
-- Product contract loading requires the frozen v1 hash even when `specs/` is absent from an installed wheel.
-- `source_verified_release_transform_registry` keeps verified visible-preserving rules instead of dropping them after promotion checks.
-- `BayesianDetectorEvidence.confirmatory_score()` refuses UNVERIFIED status. `raw_score` stays the frozen v1 field.
-- `load_bayesian_checkpoint` rejects JSON larger than 64 MiB.
-- Seed `880000` is `PUBLICLY_EXPOSED` by closed unmerged PR #98 and is not eligible as unseen validation. Added `global-seed-ledger-v1` and reserved Cycle 8 scale seeds `930000` / `940000` / `950000` before scale generation.
-- Cycle 8 scale exploratory seed `930000` U+034F x1: 0/16 then 0/32 raw transformed WM, then **1/64** on the 64-pair expansion (matched UW 0, visible projection pass, sanitizer matrix matches raw). Do not rewrite 1/64 as zero. Independent replication seed `940000` n=64 is **0/64** raw transformed WM (max score 0.557052 versus threshold 0.557099). Combined 64-pair corpora: **1/128**. Product registry fail-closes carrier sites that would violate frozen hard invariants.
-- U+034F is still not product-authorized. Do not generate `950000`. Do not inspect `830000` / `840000` / `850000`.
-- Reserved Cycle 8 density exploratory seed `960000` (`carrier density follow-up`) before generation, then scored a detector-blind U+034F space x1 versus space plus word-final letter x1 comparison. Both arms are **1/16** raw transformed WM on the same residual row. Density did not beat space x1. Not product-authorized.
-- Cycle 8 letter-x1 (U+034F after ASCII letters, visible-word invariants, quote-interior carrier policy, detector-blind selected-site cap 192) diagnostically rescored seen `930000` n=64, `940000` n=64, and `960000` n=16 to **0** raw transformed WM with clean unwatermarked controls and exact visible projection. Independent reserved seed `970000` (`intra-word carrier follow-up`) is **0/16** then **0/64**. Experimental letter-x1 **0/192** = seen `930000` n=64 plus seen `940000` n=64 plus independent `970000` n=64 (**128/192 seen**, **64/192 independent**). Space-x1 1/64 and 1/16 remain frozen.
-- Cycle 8 letter-x1 system benchmark reserved seeds `980000` and `990000` before generation. Fresh independent corpora: letter-x1 **0/128** raw transformed WM, space-x1 **1/128** on the same texts, letter max score 0.554066 versus threshold 0.557099. Local fixtures: visible projection 21/21, protected spans 21/21, Chromium `pre`/textarea/contenteditable pixel-equal on measured samples. Formal confirmation readiness is `NOT_READY`. Letter-x1 is not product-authorized and is not confirmation. Do not generate `950000`.
-- Cycle 8 letter-space margin seeds `1000000` and `1010000` were reserved before generation. Combined fresh letter-space is **1/128** on the same residual geometry; do not rewrite that as zero.
-- Cycle 8 letter-mix margin reserved seeds `1020000` and `1030000` before generation. Detector-blind U+034F/U+FE00 alternating letter mix is fresh **0/128** raw transformed WM, mix UW **0/128**, worst mix score 0.513691 versus threshold 0.557099 (gap 0.043407). Letter-x1 on those same corpora is 0/128 with max 0.527389. Chromium mix fixtures were pixel-equal on the measured `pre`/textarea/contenteditable surfaces. Mix is not product-authorized and is not confirmation. Formal confirmation readiness remains `NOT_READY`.
-- Cycle 8 letter-mix scale reserved seeds `1040000` and `1050000` before generation. Combined with `1020000`+`1030000`, mix is **0/256** raw transformed WM, mix UW **0/256**, worst mix score 0.519522 (gap 0.037577). The two-corpus 0/128 scorecard is not rewritten. This 0/256 is development scale, not confirmation.
-- Froze development mechanism `u034f-ufe00-letter-alt-v1` as `cycle8-mix-freeze-v1`. Public CLI remains empty.
-- Cycle 8 mix freeze one-shot confirmation on preregistered seeds `830000` / `840000` / `850000`: mix **0/192** raw transformed WM, mix UW **0/192**, visible **192/192**, frozen sanitizers matching raw zeros, worst mix score 0.524300 versus threshold 0.557099 (gap 0.032798). Those seeds are spent. Do not rerun looking for zero. Mix is not product-authorized. Do not generate `950000`.
+- Transforms ordinary English ASCII text. Visible words stay exactly the same.
+- In a terminal, `fuckmark` opens a paste UI. Finish with `:done`. The result is copied and not printed.
+- Pipes, quoted text, files, and `--stdin` write the payload to stdout. `--visible` prints the original visible text. `--copy` copies stream output.
+- Unsupported Unicode is returned unchanged. Empty input and invalid UTF-8 fail with an actionable error.
+- Confirmed on GPT-2 / SynthID tests: transformed text 0/192 after required sanitizers, exact visible text 192/192. This does not remove every watermark.
+- Install from a clone with `python -m pip install .`. The GitHub Release wheel is published after the immutable `v0.4.0` tag.
 
-## v0.2.0 — Exact-survival confirmation and release hardening
+## v0.3.0
 
-### Research
+Public CLI left text unchanged (no visible edits). Install and release hardening. Not the current product.
 
-- Confirmed the Cycle 4 exact post-retokenization scheduling objective on three independent frozen confirmation corpora.
-- The preregistered aggregate outcome is `CONFIRMATORY_IMPROVEMENT`.
-- Cycle-3 proxy scheduling left 8/192 transformed watermarked samples detected; exact-survival scheduling left 5/192 detected under the same inherited threshold.
-- The measured evasion rate on that frozen open Hugging Face SynthID configuration therefore changed from 95.83% to 97.40%.
-- Matched unwatermarked detections remained 2/192 in both aggregate arms.
-- Every independent 64-source confirmation corpus improved by one residual detection: 3->2, 3->2, and 2->1.
-- Added `exact-survival-greedy-key-blind-v2`, extending the frozen v1 objective with deterministic pairwise completion after single-candidate saturation.
-- Added the development-only `content-region-destruction-v1` visible candidate pool without mutating the frozen Cycle-3 ruleset hash.
-- Added raw, NFKC, Cf-strip, and combined sanitizer robustness reporting.
-- Kept U+200C explicitly quarantined as a diagnostic upper-bound mechanism and excluded it from automatic release behavior.
-- The expanded pool plus scheduler v2 reached 0/12 detected in the fresh development run, but this remains development evidence rather than a large-corpus confirmatory claim.
+## v0.2.0
 
-### Release engineering
+Historical contraction CLI. Not the current product.
 
-- Advanced the project version to 0.2.0 across package metadata, package identity, lock metadata, CLI regressions, and clean-install verification.
-- Removed the hard-coded `v0.1.0` publication fallback from `Release Engineering`.
-- GitHub Release publication now occurs only from an immutable `v*` tag.
-- Added a publication guard that requires the Git tag to exactly match the version in `pyproject.toml`.
-- Made `tools/verify_release_install.py` derive the expected version from project metadata instead of embedding a release-specific literal.
-- Upgraded the core CI and release workflows to current Node 24 GitHub Actions.
-- Retained cross-platform wheel/sdist build, `twine check`, clean-install verification, console-command verification, and SHA-256 generation.
+## v0.1.0
 
-### Documentation
-
-- Rewrote the README for the v0.2.0 release and current Cycle 4 evidence.
-- Updated all current documentation under `docs/`: CLI, installation, and release process.
-- Separated the confirmed 97.40% open-detector research result from the conservative public CLI behavior.
-- Preserved frozen historical evidence and v0.1.0 release-readiness artifacts rather than rewriting old contracts to match the new release.
-
-### Claim boundary
-
-The 97.40% figure is specific to the frozen open GPT-2 / Hugging Face SynthID Weighted Mean confirmation configuration and its fixed threshold. v0.2.0 does not claim perfect watermark removal, arbitrary-model transfer, proprietary-detector transfer, future-watermark transfer, or formal confirmation of scheduler v2.
-
-## v0.1.0 — Foundation hardening
-
-v0.1.0 established the deterministic research and release foundation used by Cycle 4. Major work included:
-
-- deterministic source/run identities, hashing, canonical serialization, and replayable evidence;
-- token alignment and exact n-gram observation geometry;
-- pinned DeepMind and Hugging Face SynthID observation adapters;
-- source-conformant Mean and Weighted Mean detector primitives;
-- conservative fixed-FPR calibration and pristine-baseline evidence;
-- content-addressed corpora, prompt provenance, generation boundaries, and matched watermarked/control validation;
-- immutable protected-span extraction and hard semantic invariants;
-- deterministic transform registries, candidate enumeration, conflict graphs, traces, and replay;
-- TinyDev and MidDev experiment infrastructure with separate planning and scoring paths;
-- release-safe CLI behavior for interactive, stream, file, clipboard, and atomic-output use;
-- wheel and source-distribution verification on Linux, macOS, and Windows;
-- MIT licensing, package metadata, project URLs, and tagged GitHub Release support;
-- explicit rejection records for experiments that failed preregistered effectiveness gates;
-- quarantining of invisible-format-character experiments that failed sanitizer durability.
-
-The detailed v0.1.0 research chronology remains preserved in the repository history and frozen artifacts under `specs/`. Those historical contracts are not rewritten by later releases.
+First tagged release.