docs: retire the warning sign from the five shipped operator docs (BACKLOG #1265) - #952

Retires the warning sign (U+26A0) from the five shipped operator docs, which BACKLOG #1265 names
as its defensible first slice. The item stays open: the two ledger files and the ADR decision are
untouched and still sequenced last.

What changed

17 sites in SECURITY.md, PHI.md, INSTALL-GUIDE.md, DEPLOYMENT.md and CONNECTIONS.md now
say the word the sentence means. Not a mechanical substitution, so the words differ by meaning:

The 2 cross-references are the ones worth reading. CONNECTIONS.md said "see the [glyph]
above" and "the whole point of the [glyph] above", the second one 60 lines from its target.
Neither told a reader which mark was meant. They now name the [inbound]-versus-inbound(...)
distinction they point at. That is the exact scope defect CLAUDE.md section 11 describes, so these
two are the argument for the sweep rather than a side effect of it.

CLAUDE.md section 11 now carries a fresh census. The BACKLOG item gets a dated amendment.

The three required proofs

1. Census, ledger counts as the positive control. Counted over git-tracked files decoded as
UTF-8, at 172b1327c.

 before after
docs/SECURITY.md 5 0
docs/PHI.md 2 0
docs/INSTALL-GUIDE.md 1 0
docs/DEPLOYMENT.md 2 0
docs/CONNECTIONS.md 7 0
SLICE TOTAL 17 0
REPO-WIDE 496 / 75 files 479 / 70 files
positive control docs/BACKLOG.md 127, BACKLOG-CLOSED.md 93 (both passes)

The slice reaches zero while the control still reports non-zero on the same pass, so the zeros
are the files and not a dead pattern.

2. The grep escape was avoided rather than survived. No shell pattern was used. The census and
the test both build the codepoint in Python (chr(0x26A0), "\N{WARNING SIGN}"), so no non-ASCII
byte and no \u escape ever crossed a shell. The instrument was still checked against the figure
everyone agrees on before any zero was trusted: it reproduces docs/BACKLOG.md at 127 and
BACKLOG-CLOSED.md at 93.

3. backlog_status_check.py still passes at its full item count. Identical before and after
the amendment: OK - 679 backlog items, each declaring exactly one status, scanning
docs/BACKLOG.md (442) and BACKLOG-CLOSED.md (237), exit 0. The one advisory warning is
pre-existing (#1426, unrelated). parse_items reports #1265 still open, closed=[] - exactly
one status banner.

Two rows of the filed census table are wrong, and both are left standing

Corrected in CLAUDE.md section 11, not edited in the ledger table, because the wrong version is
what a later reader would otherwise re-derive. Both are SDS-3.8, an answer to an adjacent question:

• The web console is not zero. The table reads it as zero by counting packaging/. The
  console's source is messagefoundry_webconsole/ (36 tracked files), which carries 2 -
  _external.py and routes/oidc.py. The item's prose claim "the web console has NONE" is
  false at this ref.
• There is no harness/ row at all, so 26 occurrences across 4 files sat outside every
  bucket. The filed buckets sum to 464 against a printed 496; the corrected ones sum to 479.

My file count (75) differs from both the filed 80 and the re-scored 81. I did not chase the
difference: my denominator is stated (git-tracked, UTF-8 decodable) and the buckets sum to the
total, which the filed ones do not.

Decisions the brief left open

1. Added a regression guard, tests/test_operator_docs_no_warning_sign.py, and registered it
   in two places the brief did not mention. Both were necessary rather than optional:
   tests/tooling_manifest.txt (without the entry test_tooling_partition.py fails, reddening
   three required legs - measured) and ci.yml DOC_GUARDS (the guard's whole subject is
   Markdown, so the code == 'true' gated pytest legs skip the only change class that can break
   it - the docs-only blind spot the step's own comment records twice).
2. The positive control is planted in the test, not borrowed from docs/BACKLOG.md. Borrowing
   the ledger's 127 would make this guard red for the wrong reason the day the last slice of #1265
   lands.
3. The guard also refuses an orphan U+FE0F. Every U+26A0 in this corpus trailed one, and a
   stray variation selector is invisible in review.
4. _SWEPT holds repo-relative paths, not bare names, so a later slice extends it by adding a
   line. What remains is not all under docs/.
5. Paragraphs were not reflowed. Two replacements left a line slightly long. There is no
   markdown line-length lint, and at least four other live sessions have these files open, so a
   minimal diff is worth more than a tidy wrap.
6. Did not build a census baseline file. See the open question below.

Checks

Red-first counted by case, not by "something went red". Reverting the five docs to their
pre-sweep content and running the new module gave 10 failed, 1 passed: the 5 glyph cases and
the 5 orphan-selector cases, each naming its file, with the positive control the single pass. The
control passing is what makes the 10 reds mean the files rather than a broken scanner.

The full suite did not finish, and I am not claiming it passed. Two runs against the committed
tree were killed by the host for memory pressure, not by a test - this box is running a dozen or
more concurrent sessions. The first reached 57% with 3156 passed, 45 skipped, 0 F, 0 E
(counted off the progress characters, not inferred from the absence of a summary line); the second
was killed at 4% with the same shape. That is partial evidence, not a pass. I stopped after two
attempts rather than grind.

What makes this a low-risk gap rather than an unknown: the change touches no engine code. It is
five Markdown files, two ledger/record documents, one new test, one manifest line and one workflow
line. Every test that can read those files was run to completion and named above.

Legs a Builder cannot see, which must be read on the runner: windows-service-smoke (NSSM),
and the docs-only lane itself - this PR changes DOC_GUARDS, so the doc-guards step's membership
and the tooling job's manifest gate should both be confirmed green there.

Concurrency

Ten or more live sessions have these files open. tooling_manifest.txt had uncommitted changes
in another worktree when I edited it; I compared the entry sets first - that session adds
test_verdict_divergence_advisory.py 24 lines away from my insertion, so the two additions merge
cleanly. SECURITY.md, PHI.md, CONNECTIONS.md, BACKLOG.md and ci.yml all carry committed
changes on other branches. My edits are single-line and glyph-scoped, but a rebase is likely.

Merge safety

main moved one commit under this branch (ebdfa44a6), touching docs/BACKLOG.md and
docs/CONNECTIONS.md. git merge-tree --write-tree origin/main HEAD exits 0: both auto-merge,
no conflict. A clean merge is not the question that matters though, so the merged tree was
censused directly - all five swept files are still 0, and the control still reads 127. The PR
shows BEHIND, not CONFLICTING.

Two things found in passing, outside this slice

Named rather than numbered, since I have allocated neither.

1. The banner alphabet is used decoratively outside the two sanctioned files. Measured at this
   ref: docs/CONNECTIONS.md carries 124 U+2705, 18 U+274C and 3 U+23F3 in its connector-parity
   table, and docs/PHI.md carries 2 U+2705. CLAUDE.md section 11 says nothing outside
   BACKLOG.md and BACKLOG-CLOSED.md may adopt that vocabulary. That is a different
   population from #1265 and wants its own item. The new test's docstring says so explicitly, so
   nobody reads a green run here as covering it.
2. Nothing stops the unswept population growing. The item's own re-score measured it drifting
   up, 496 to 499 in six days with nobody sweeping. A prose number in two documents, verified by
   nothing, is the instrument that already failed here twice. The repo's idiom for this is a
   checked-in baseline the guard re-derives and ratchets downward
   (scripts/quality/username_access_key_baseline.txt, scripts/docs/asvs_tally_baseline.txt).
   Deliberately not built here - it is machinery for the whole population, and this brief scoped me
   to one slice.

Generated with Claude Code

https://claude.ai/code/session_01Mk1rsUuxp823qZs4tecMd6