docs: Phase 7 — language tags, pagelint --fix, and the acceptance pass (spec 011 complete)

[iancooper]

Closes out Spec 011. Three commits, one per task; 43 of 43 tasks done.

Task 7.2 — language tags on the last 34 fences, rule 4 to a repo-wide error

The backfill and the rule that stops it recurring, in one commit — tagging without
enforcing leaves the tags to decay like every other unenforced convention here, which
is what the audit measured in the first place.

Not one of the 34 was code: 1 bash block (two dotnet add package lines) and 33
text — ASCII flow diagrams, directory trees, validation-message dumps, trace output.
That is the explanation for the debt rather than a curiosity. An author writing C#
reaches for ```csharp because they want the highlighting; an author drawing a
box-and-arrow diagram has no language in mind and types a bare fence. The untagged
population was never latent code debt — it was the absence of a name for "not code".

Corroborated by a number that did not move: the using-directive debt is 802 blocks
across 93 pages before and after. Tagging any of the 34 csharp would have added it to
rule 6's population.

pagelint.py goes 836 warnings → 802, and 802 is the using debt entire. Rule 4 was
proved red before it was trusted green: retagging a fence back to bare gives
1 errors and exit 1.

Task 7.3 — pagelint.py --fix

A version bump is 110 banners, and retyping one segment 110 times gets abandoned
half-done. Rehearsed end to end with APPLIES_TO temporarily at V11/V5: 110 stale,
110 fixed, 0 left for a human. Page types survived exactly (50/33/27), all five
Prerequisites segments survived — checked first, because that is the failure
apply_banners.py actually shipped — one line changed per page, and the 17 files
without a trailing newline still have none. Second run is a no-op.

There is no migration map. A stale value names a set of products and the fix is
whichever APPLIES_TO entry names exactly that set, so one edit to the tuple is the
whole bump.

It never decides a page type. Given **Guide** · Applies to **Brighter V9** it
fixes the version, leaves the type, and the page still fails rule 2 — --fix cannot
launder a bad page type into a green build.

The language-tag half had ground truth: the 34 fences 7.2 tagged by hand. Replayed
against that tree it tags 26 and holds 8, and all 26 match the hand verdict. No
wrong tag; the only failure mode is doing nothing and saying so.

Task 7.4 — acceptance pass, AC1–AC8

Seven passed as they stood. AC5 failed and is fixed: the linter can emit eight rule
labels and the ledger listed seven — NO H1 was never written down. Nothing was broken
by it, because every page has an H1 and the rule has never fired, which is exactly why
it survived four sessions of the ledger being read and edited. A rule that never fires
is invisible to everything but an enumeration.

A second drift of the same shape turned up while checking AC6: docs.yml's comment
still said untagged fences stay warnings.

Two checks that had never been run, both of which could have been false silently: every
page's banner type matches its reviewed verdict in pagetypes.tsv (110/110), and every
page's Applies to matches the TSV's applies column (110/110, 100/5/5). AC4 was
re-derived without the linter — 688 distinct non-navigation ## slugs, 0 repeated.

State

pagelint.py 0 errors / 802 warnings across 110 pages; linkcheck.py clean at 112
files; --changed origin/master green and non-vacuous. The remaining 802-block
using-directive debt is deliberate and is AC1's baseline to shrink.

🤖 Generated with Claude Code

https://claude.ai/code/session_012tcdwxVb8NmKaX2S6fvyFg