Wave 6: decision-council surface fixes, standing every-test-wired check, locale-stable gates, audit-pr footer markup

.claude/skills/communication-structure/INDEX.md

@@ -6,7 +6,7 @@ Generated by `scripts/skill_index.py` (do not hand-edit; rerun it after changing
 
 | File | Read this when... | Tok |
 |---|---|---|
-| `references/artifact-templates.md` | you draft a status update, PR body, release note, decision memo or executive summary, incident summ… | 2292 |
+| `references/artifact-templates.md` | you draft a status update, PR body, release note, decision memo or executive summary, incident summ… | 2308 |
 | `references/leadership-status-update.md` | you write any status update a non-technical leader will read: a tracker project update, an issue st… | 500 |
 
 ## Scripts

.claude/skills/communication-structure/references/artifact-templates.md

@@ -41,8 +41,8 @@ verified and command. Fenced code does not count toward the budget.
 `audit-pr` is the body of an audit, security or hardening PR. `sections` fails, at any length, when a required heading
 (every heading of the template below except `Dependency upgrades`) is missing, `What changes for reviewers and users` is
 empty (items, or `None`), or `Not fixed` carries a file:line, route, URL or file path (counts only). A closing footer
-(`---` or `🤖 Generated`, then only blank, `Generated with` and `Co-Authored-By:` lines) is not checked; other text after
-a `---` is. With `--range BASE..HEAD` it also fails when a changed file is not named under `Changes by area`, by full path
+(`---` or `🤖 Generated`, then only blank, `Generated with` and `Co-Authored-By:` lines, any of them in `<sub>`, `<sup>`,
+`<small>`, `*` or `_` markup) is not checked; other text after a `---` is. With `--range BASE..HEAD` it also fails when a changed file is not named under `Changes by area`, by full path
 or by a containing directory with a trailing slash. The guard hook picks this kind when the `gh pr create|edit` title says audit, security
 or harden, and for `create` passes the diff against `--base` (default `origin/HEAD`).
 

.claude/skills/communication-structure/scripts/comms_lint.py

@@ -20,8 +20,9 @@
 AUDIT_SECTIONS heading is present, "What changes for reviewers and users" is not empty (items, or "None"), "Not fixed"
 holds counts only (no file:line, route or URL), and, with --range BASE..HEAD, every file the range changes is named
 under "Changes by area" by its full path or by a containing directory with a trailing slash. A trailing attribution footer
-(a `---` rule or `🤖 Generated` line, then only blank, `---`, `Generated with/by` and `Co-Authored-By:` lines) is not part
-of the last section; any other text after a `---` is still checked.
+(a `---` rule or `🤖 Generated` line, then only blank, `---`, `Generated with/by` and `Co-Authored-By:` lines; those lines
+may sit in `<sub>`, `<sup>`, `<small>`, `*` or `_` markup) is not part of the last section; any other text after a `---` is
+still checked.
 
 Prints `FAIL <rule> L<n>: <problem>. Fix: <exact fix>` per failure and a score line; exit 1 when any non-advisory
 rule fails (below the bar), 0 otherwise, 2 on usage, unreadable input or a --range git cannot resolve (fail closed).
@@ -239,8 +240,10 @@ def _meeting(res, prose):
 DETAIL = re.compile(r"\.[A-Za-z]\w{0,7}:\d+|#L\d+|\blines?\s+\d+|\b(?:GET|POST|PUT|PATCH|DELETE|HEAD|OPTIONS)\s+/|"
                     r"(?<![\w/:.-])/[\w.-]+|\bhttps?://|\b[\w.-]{1,100}/[\w./-]{0,200}\.[A-Za-z]\w{0,7}\b")
 # one attribution-footer line, matched whole: a `---` rule or `🤖 Generated ...` (either starts a footer), `Generated with/by ...`,
-# `Co-Authored-By: ...` or nothing. ponytail: a line is judged by its prefix, so detail on a `Generated ...` or `Co-Authored-By:` line goes unchecked; tighten the shapes if that is abused
-FOOTER = re.compile(r"(?:(?P<mark>-{3,}|🤖\s*Generated\b.*)|Generated (?:with|by)\b.*|Co-Authored-By:.*)?", re.I)
+# `Co-Authored-By: ...` or nothing. The `🤖 Generated`, `Generated with/by` and `Co-Authored-By:` shapes may open in <sub>, <sup>, <small>, `*` or `_` markup (the closer is part of the tail).
+# ponytail: a line is judged by its prefix, so detail on a `Generated ...` or `Co-Authored-By:` line goes unchecked; tighten the shapes if that is abused
+_MARKUP = r"(?:</?(?:sub|sup|small)>|[*_]){0,6}"  # bounded: an unbounded repeat would be quadratic on a long run of `*`
+FOOTER = re.compile(r"(?:(?P<mark>-{3,}|" + _MARKUP + r"🤖[\s*_]*Generated\b.*)|" + _MARKUP + r"Generated (?:with|by)\b.*|" + _MARKUP + r"Co-Authored-By:.*)?", re.I)
 
 
 def _name(text):
@@ -265,7 +268,7 @@ def _sections(text):
     runs to the next heading of the same or a higher level (sub-headings stay inside it); the first of a repeated heading
     wins; HTML comments are dropped, so a template hint is not content. A trailing attribution footer is not content either:
     the last section ends at the first `---` rule or `🤖 Generated` line that is followed only by FOOTER-shaped lines (blank,
-    `---`, `Generated with/by ...`, `Co-Authored-By: ...`) to the end; any other text after a rule keeps the rule as content."""
+    `---`, `Generated with/by ...`, `Co-Authored-By: ...`, in markup or not) to the end; any other text after a rule keeps the rule as content."""
     lines = _blank_comments(text).splitlines()
     heads, fence, foot = [], False, len(lines)
     for i, ln in enumerate(lines):

.claude/skills/idea-critic/INDEX.md

@@ -7,7 +7,7 @@ Generated by `scripts/skill_index.py` (do not hand-edit; rerun it after changing
 | File | Read this when... | Tok |
 |---|---|---|
 | `references/anti-rationalization.md` | you are about to emit PASS_TO_USER, skip a hat, or accept a shortcut (see SKILL.md, Escalate). | 332 |
-| `references/council.md` | a decision is high-stakes AND hard to reverse, and one inline pass is too weak: you want independen… | 2267 |
+| `references/council.md` | a decision is high-stakes AND hard to reverse, and one inline pass is too weak: you want independen… | 2379 |
 | `references/decision-hygiene.md` | the owner faces a hard, high-stakes call of their *own* — a pivot, quitting or killing a line of wo… | 1137 |
 | `references/full-map.md` | the quick attack in SKILL.md is not enough: an owner-decision or high-blast claim, an agent-origina… | 3395 |
 

.claude/skills/idea-critic/references/council.md

@@ -23,7 +23,7 @@ neither and it convenes (legacy behavior), so declare them.
 | The evil twin argues the strongest dissent | `--twin`: after the decision, one call (`lenses/evil-twin.txt`) writes a pre-mortem and the strongest case against it; kept in `minority_report` with the dissenting seats' names (their answers stay verbatim in `dissent`) |
 | Pre-mortem | every seat answers `premortem`, `kill_criterion`, `leading_metric` (recorded in `premortems`), then the twin's |
 | Deterministic aggregate; the chair breaks ties | decision = highest weighted mean of seat scores (0-5 per option and criterion); a tie goes to `--chair`'s own top option, then to option order (`do-nothing` first). The verdict tally ties go to the chair, else `HOLD` > `REVISE` > `PASS_TO_USER` |
-| Output: decision, kill criteria, metric, review date | record keys `decision`, `kill_criteria`, `leading_metric`, `review_date` (`--review-date`, else today + `--review-days`, default 30). A closer no seat supplied is `UNVERIFIED`, never invented |
+| Output: decision, kill criteria, metric, review date | record keys `decision`, `kill_criteria`, `kill_criteria_by_lens` (`{lens: text}`), `leading_metric`, `review_date` (`--review-date`, else today + `--review-days`, default 30). A closer no seat supplied is `UNVERIFIED`, never invented |
 | Convene only when it matters | the gate above |
 
 ## Run
@@ -36,7 +36,7 @@ python3 scripts/veche.py --cmd "claude -p --strict-mcp-config" \
   --stakes high --reversibility hard
 ```
 
-- `--cmd` reads the prompt on stdin and prints the answer. Without `--lenses` it is solo-3: one cmd, the three hats.
+- `--cmd` reads the prompt on stdin and prints the answer. A custom `--cmd` reply needs a JSON object with a `verdict` (`HOLD|REVISE|PASS_TO_USER`) and string-only `claims` items; prose and stray braces around it are skipped, the first object with a valid `verdict` wins, and with `--option` the `scores` map may sit under `scores` in that object, or in any other object (bare or under `scores`). Without `--lenses` it is solo-3: one cmd, the three hats.
 - `--option` names must be short tokens (letters, digits, `-`, `_`, starting with a letter or digit; at most 32): seats echo them as JSON keys. Others fail before any call.
 - The packet is shown with numbered lines and its path under `--root`, so a model with no file tools
   can still cite `path:line`; keep the packet under `--root`.
@@ -56,7 +56,7 @@ python3 scripts/veche.py --cmd "claude -p --strict-mcp-config" \
   checked before each call. Any cap hit, non-zero exit, timeout or unparseable answer yields
   `COULD_NOT_DECIDE` (exit 3), never a guessed verdict, with `dropped` reasons and `cost`. A timed-out command's grandchildren may outlive
   it, so use a cmd that exits cleanly. `--dry-run` prints planned calls.
-- `--json-out PATH` also writes the record there (a bad path fails before any call). A gate skip prints `NOT_CONVENED` on stderr, noting an ignored `--twin`.
+- `--json-out PATH` also writes the record there (a bad path fails before any call; a write that fails at the end exits 2 only when the verdict would exit 0, so 3 and 4 stay). A gate skip prints `NOT_CONVENED` on stderr, noting an ignored `--twin`.
 - Validate: `scripts/validate_verdict.py --file record.json` accepts the record (it has `tally`),
   `COULD_NOT_DECIDE` and `NOT_CONVENED` with a `reason`, and a record with a `decision` only when it also
   carries `kill_criteria`, `leading_metric` and an ISO `review_date`.
@@ -80,8 +80,8 @@ which yields the `default_action` above. A malformed `--criteria` error states t
 When the critique round runs, each seat also ranks the other (anonymized, lens names masked) answers best
 first; valid rankings are Borda-aggregated into `peer_ranking` (lens to mean points, informational: the
 verdict still comes from the tally and evidenced attacks). There is no separate adapter: any other council
-can be a seat or the whole panel through the existing `--cmd`, a script that reads the prompt on stdin and
-prints the JSON contract. A seat that fails (cap, timeout, cmd exit) is dropped and lowers confidence; a
+can be a seat or the whole panel through the existing `--cmd` (a script; its reply contract is under Run).
+A seat that fails (cap, timeout, cmd exit) is dropped and lowers confidence; a
 failed critique call or twin keeps the round-1 verdict. Versus the fetched sources (`docs/standards-index.md`):
 
 | | llm-council | Du et al. debate | Mixture-of-Agents | Veche |

.claude/skills/idea-critic/scripts/veche.py

@@ -20,18 +20,19 @@
 dropped with their reasons and the call cost, and a stderr COUNCIL_BLOCKING_WORK alarm (fired while the council
 runs, once --alarm-s has passed), so a stuck council never stalls unrelated work. --cmd can wrap any
 external council (a script reading the prompt on stdin, printing the JSON contract).
-Side effects: spawns the given commands, reads files under --root, writes --json-out (an unwritable path exits 2, also after the
-record was printed). Exit 0 decided.
+Side effects: spawns the given commands, reads files under --root, writes --json-out (a temp file beside it, then os.replace; an unwritable path exits 2,
+also after the record was printed, unless the verdict already exits 3 or 4). Exit 0 decided.
 """
 from __future__ import annotations
 
-import argparse, concurrent.futures, datetime, threading, json, os, random, re, secrets, shlex, subprocess, sys, time
+import argparse, concurrent.futures, contextlib, datetime, threading, json, os, random, re, secrets, shlex, subprocess, sys, time
 from pathlib import Path
 
 VERDICTS = ("HOLD", "REVISE", "PASS_TO_USER")  # most cautious first: ties go left
 HATS = "skeptic,better-way,kill-criteria"
 NAME = re.compile(r"^[a-z][a-z-]{0,30}$")
 TOKEN = re.compile(r"[A-Za-z0-9][A-Za-z0-9_-]{0,31}")  # an --option name: every seat must echo it exactly as a JSON key
+TOKEN_RULE = f"letters, digits, - or _, starting with a letter or digit, at most 32 chars: {TOKEN.pattern}"
 CAP = "cap reached before call"
 CITE = re.compile(r"^(?P<p>[^\s:][^:]*):(?P<a>\d+)(?:-(?P<b>\d+))?$")
 DO_NOTHING = "do-nothing"
@@ -72,10 +73,11 @@ def norm(t: str) -> str:
 
 def uniq(items: list) -> list:
     """Drop near-duplicate strings, keeping the first wording and the order: case, spacing and punctuation are
-    ignored; digits and comparison operators are not (5% and 6% stay two items, and so do `< 200 ms` and `> 200 ms`)."""
+    ignored; digits and comparison operators are not (5% and 6% stay two items, and so do `< 200 ms` and `> 200 ms`); `≤` and `≥`
+    are the same operators as `<=` and `>=`, so the Unicode and ASCII spellings fold together."""
     seen: dict = {}
     for t in items:
-        seen.setdefault(" ".join(re.findall(r"\w+|[<>≤≥]=?", t.lower())), t)
+        seen.setdefault(" ".join(re.findall(r"\w+|[<>]=?", t.lower().replace("≤", "<=").replace("≥", ">="))), t)
     return list(seen.values())
 
 
@@ -200,7 +202,7 @@ def run(a, st=None) -> dict:
     names = [o.strip() for o in (a.option or []) if o.strip() and o.strip() != DO_NOTHING]
     for o in names:
         if not TOKEN.fullmatch(o):
-            raise Stop(f"bad --option {o!r}: use a short token (letters, digits, - or _, at most 32 chars), e.g. ship; "
+            raise Stop(f"bad --option {o!r}: use a short token ({TOKEN_RULE}), e.g. ship; "
                        "every seat must repeat it exactly as a JSON key")
     options = [DO_NOTHING] + names
     if len(options) > 8 or len(set(options)) != len(options):
@@ -388,6 +390,7 @@ def others_of(i):
         ["UNVERIFIED: no seat wrote a pre-mortem"]
     kills = uniq([x["kill"] for x in answers if x["kill"]] + ([twin["kill_criterion"]] if isinstance(twin.get("kill_criterion"), str) and twin["kill_criterion"].strip() else []))
     rec["kill_criteria"] = kills or ["UNVERIFIED: no seat stated a measurable tripwire"]
+    rec["kill_criteria_by_lens"] = {x["lens"]: x["kill"] for x in answers if x["kill"]}  # each seat's own wording, before the dedupe
     metrics = {x["lens"]: x["metric"] for x in answers if x["metric"]}
     rec["leading_metric"] = metrics.get(a.chair) or next(iter(metrics.values()), "UNVERIFIED: no seat named a metric")
     rec["minority_report"] = {"twin": twin or None, "dissenting_seats": [x["lens"] for x in dissent]}  # their answers are in `dissent`
@@ -424,7 +427,7 @@ def stop():
 
 def main(argv=None) -> int:
     """CLI entry. Prints the JSON record (and writes it to --json-out). Exit 0 decided, 2 --json-out unwritable (the record is still
-    printed), 3 COULD_NOT_DECIDE (fail closed), 4 NOT_CONVENED."""
+    printed; a write that fails at the end changes only exit 0, never 3 or 4), 3 COULD_NOT_DECIDE (fail closed), 4 NOT_CONVENED."""
     ap = argparse.ArgumentParser(description=__doc__.split("\n")[0])
     ap.add_argument("--cmd", action="append", required=True)
     ap.add_argument("--claim", required=True)
@@ -433,12 +436,15 @@ def main(argv=None) -> int:
     ap.add_argument("--origin", default="agent-originated", choices=("owner-request", "agent-originated"))
     ap.add_argument("--root", default=".")
     ap.add_argument("--lenses", default=HATS, help="comma list of lens files; seats: cpo,cto,cfo,ciso,user-advocate")
-    ap.add_argument("--option", action="append", help="an option to score: a short token such as ship (repeat; do-nothing is always added; include the claim)")
+    ap.add_argument("--option", action="append", help=f"an option to score, a short token ({TOKEN_RULE}) since every seat repeats it as a JSON key, e.g. ship (repeat the flag; do-nothing is always added; include the claim)")
     ap.add_argument("--criteria", help="name:weight,... fixed before scoring; required with --option")
     ap.add_argument("--chair", help="a lens name: breaks ties in the verdict tally and the option scores")
     ap.add_argument("--twin", action="store_true", help="add the evil twin: pre-mortem and strongest dissent")
-    ap.add_argument("--stakes", choices=("low", "medium", "high"))
-    ap.add_argument("--reversibility", choices=("easy", "hard"))
+    ap.add_argument("--stakes", choices=("low", "medium", "high"),
+                    help="low or medium skips the council (NOT_CONVENED, exit 4, no seat called) only if --reversibility easy is also given; high, or no --reversibility, convenes")
+    ap.add_argument("--reversibility", choices=("easy", "hard"),
+                    help="easy skips the council (NOT_CONVENED, exit 4, no seat called) only if --stakes low or medium is also given; hard, or no --stakes, convenes. "
+                         "easy also makes the default_action of a COULD_NOT_DECIDE record proceed, else hold")
     ap.add_argument("--review-date", help="ISO date to re-check the decision (default: today + --review-days)")
     ap.add_argument("--review-days", type=int, default=30)
     ap.add_argument("--max-calls", type=int, default=12)
@@ -447,15 +453,28 @@ def main(argv=None) -> int:
     ap.add_argument("--timeout", type=int, default=120)
     ap.add_argument("--alarm-s", type=int, default=300, help="print COUNCIL_BLOCKING_WORK to stderr once the council has run this long, while it runs")
     ap.add_argument("--progress-s", type=int, default=30, help="stderr heartbeat every N seconds while the council runs (0 = off)")
-    ap.add_argument("--json-out", help="also write the JSON record to this file (opened before any call, so a bad path fails early)")
+    ap.add_argument("--json-out", help="also write the JSON record to this file on every outcome, via a temp file in the same directory and os.replace; "
+                    "an empty file is written first, before any call, so a bad path fails early. The directory must be writable: a read-only file in it is replaced, "
+                    "while a read-only directory or a bind-mounted file (os.replace cannot replace a mount point) exits 2")
     ap.add_argument("--seed", type=int, default=0)
     ap.add_argument("--dry-run", action="store_true")
     a = ap.parse_args(argv)
 
-    def write_out(text):  # an unwritable --json-out is a usage error (exit 2, no traceback), at the start and at the end
+    def write_out(text):  # atomic: a reader sees the old file or the whole new one. ponytail: no fsync, so not durable across power loss
+        real = os.path.realpath(a.json_out)  # a symlinked path keeps pointing at its target
+        tmp = f"{real}.{secrets.token_hex(4)}.tmp"
         try:
-            Path(a.json_out).write_text(text)
-        except OSError as e:
+            if os.path.exists(a.json_out) and not os.path.isfile(a.json_out):  # a device, FIFO or directory: renaming over it would replace the node itself
+                Path(a.json_out).write_text(text)  # (a directory fails here, as before)
+            else:
+                with open(tmp, "x", encoding="utf-8") as f:
+                    f.write(text)
+                with contextlib.suppress(OSError):  # an existing record keeps the mode its owner gave it
+                    os.chmod(tmp, os.stat(real).st_mode & 0o7777)
+                os.replace(tmp, real)
+        except OSError as e:  # an unwritable --json-out is a usage error (exit 2, no traceback), at the start and at the end
+            with contextlib.suppress(OSError):
+                os.unlink(tmp)
             ap.error(f"--json-out: cannot write {a.json_out}: {e.strerror}")
 
     if a.json_out:
@@ -487,9 +506,13 @@ def write_out(text):  # an unwritable --json-out is a usage error (exit 2, no tr
                   f"default_action={rec.get('default_action', 'n/a')}", file=sys.stderr)
     out = json.dumps(rec, indent=2)
     print(out)
+    code = {"COULD_NOT_DECIDE": 3, "NOT_CONVENED": 4}.get(rec["verdict"], 0)
     if a.json_out:
-        write_out(out + "\n")  # the record is already on stdout, so an unwritable path loses nothing
-    return {"COULD_NOT_DECIDE": 3, "NOT_CONVENED": 4}.get(rec["verdict"], 0)
+        try:
+            write_out(out + "\n")  # the record is already on stdout, so an unwritable path loses nothing
+        except SystemExit as e:  # the message is printed; a blocking verdict exit (3, 4) must not be turned into 2
+            code = code or e.code
+    return code
 
 
 if __name__ == "__main__":