docs: tag bare BrightScript code fences + add brightscript-fence-required lint rule

[bblietz]

Summary

Audit step 2 from dev-doc-review-report-2026-05-16.md §16. Tags every bare ``` fence whose body is confidently BrightScript so the renderer can apply syntax highlighting, AND ships a docs-lint rule (`brightscript-fence-required`) to prevent regression.

The new classifier is the same module used by both the one-shot sweep tool and the new lint rule, so the rule's notion of "looks like BrightScript" is exactly what produced this PR's tags.

Tag breakdown

568 fences auto-tagged across 154 files:

Language	Count	Notes
brightscript	499	315 strong-signal, 143 directory-context, 8 hand-patched (deep-indented / blockquote fences missed by sweep regex; lint rule caught them via mdast)
c	25	Debugger-protocol struct/enum in socket-based-debugger.md. Tagging as brightscript would actively miscolor them.
text	22	Roku channel manifest snippets, BRS debug dumps (<Component: roAssociativeArray> = { ... }), URL templates
json	7
http	6
xml	1

76 fences intentionally left bare. They're Robot Framework test files (*** Settings ***), single-line pkg:/... paths, RAF API schema descriptions (adPods : [{viewed : Boolean, ...}]), math notation (M = C(-1) S R C T), release-notes literal dumps (a from ' {...}). Leaving them bare prevents the renderer from trying to highlight them as something they aren't.

New files

• .github/scripts/docs-lint/lib/code-fence-classifier.mjs -- shared classifier returning {lang, confidence, signals}. Conservative; prefers unknown over guessing. Used by both the one-shot sweep tool AND the lint rule below.
• .github/scripts/docs-lint/rules/code-fences.mjs -- new brightscript-fence-required lint rule: any bare fence whose body classifies as BrightScript with confidence >= 0.7 is reported as an error. Wired into index.mjs RULES array.

How the classifier works

BrightScript signals (any one high-spec or two weak signals -> tag as brightscript):

• Statement keywords: sub/function/end sub/end function/endif/endwhile/endfor
• as <Type> annotations (String, Integer, Boolean, Object, Dynamic, Float, ...)
• CreateObject(\"ro...\") -- case-insensitive on the Ro prefix
• m.top/m.global/m.video and any m.field reference, assignment, or method call
• ro-prefixed object names (roSGNode, roMessagePort, roByteArray, ...)
• BRS REPL prompt (BrightScript> )
• BRS line comments (')
• Hex literals (&h...)
• Library \"...brs\" statement
• Roku_Ads() framework call
• observeFieldScoped / callFunc SceneGraph node methods

Non-BrightScript filters (decisive -- prevents BRS misclassification):

• XML opener (<?xml, SceneGraph component tags with attribute/close-bracket boundary -- explicitly NOT matching <Component: ...> BRS debug-dump syntax)
• C struct/enum declarations + C int types (debugger protocol)
• Shebang or shell prompts
• HTTP request lines (GET /path OR POST https://...)
• Roku channel manifest key=value lines (title=, mm_icon_focus_hd=, requires_*_drm=, ...)
• Single URL line / long base64-like token
• HLS m3u8 directives (#EXT-X-...)
• Single CamelCase identifier per line (ECP button names like ChannelUp / ChannelDown)
• Robot Framework markers (*** Settings ***, *** Variables ***, ...)
• RAF schema descriptions (key : Type shape definitions)
• Component callback output dumps (init(): v2 style)

Directory-context pass (sweep tool only -- the lint rule is path-blind):

In confirmed BRS directories (REFERENCES/brightscript/, REFERENCES/scenegraph/, DEVELOPER/core-concepts/, DEVELOPER/advertising/, DEVELOPER/media-playback/, DEVELOPER/getting-started/, DEVELOPER/dev-tools/, DEVELOPER/discovery/, DEVELOPER/voice/, DEVELOPER/performance-guide/), fences that don't hit strong classifier signals AND don't contain explicit non-BRS content (above) are promoted to brightscript with directory-context justification. Catches 1-line BRS method calls like adIface.setAdUrl(myAdUrl) that have no distinctive BRS keyword on their own but live in unambiguously-BRS docs.

Test plan

• node .github/scripts/docs-lint/index.mjs shows ZERO brightscript-fence-required errors after the sweep
• Lint baseline preserved: same 3 pre-existing findings in REFERENCES/brightscript/interfaces/ifscreen.md (1 pipe-no-blank-above error + 2 html-blank-between-tags warnings)
• Dry-run sample of 25 dir-context promotions reviewed; 0 false positives in final classifier config
• All 8 lint-flagged fences my sweep regex missed (blockquote-prefixed > \``and deep-indented ```` inside list items) hand-patched and re-verified
• CI run on this PR

Out of scope

• Audit's roku-pay-best_practices underscore mismatch claim -- already fixed upstream; not visible in current v2.0.
• One pre-existing malformed fence in scenegraph control-nodes docs (createobject(\"roprogramguide\")\``` -- info string has code in it, looks like a missing close fence). Flagged for manual review; not in scope here.
• The 76 truly-ambiguous fences left bare (Robot Framework, schemas, paths, release-notes literals).

Related

Part of the May 16 audit cleanup plan. Step 1 (typos + _order.yaml + empty-link) is in flight as PR #52. Next sweep PRs in the queue: terminology canonicalization (~25 hits), broken-absolute-link sweep (252 remaining).