feat(tag,help): lowercase-english-only --tag + colorized help

[alex-mextner]

Implements two ROADMAP items (CTO 2026-06-16): "tg --tag: lowercase-english only" and the tg side of "ONE shared help-formatter" (match the ecosystem color scheme until the Python migration lands the real shared lib).

1. --tag is lowercase-english ONLY

Accepted values: answer / decision / problem / report. Russian aliases (ОТВЕТ/РЕШЕНИЕ/ПРОБЛЕМА/ОТЧЁТ) and uppercase / mixed-case / unknown words are now rejected at parse time with a 3-part error and a non-zero exit, replacing the old soft-render-and-warn:

$ tg --tag ANSWER --reply-to 1 "x"
Error: invalid --tag 'ANSWER': tags must be lowercase english. Use one of: answer, decision, problem, report
$ echo $?
1

• validateTag() (features/render/tag.ts) is the gate, called in parseArgs so every send path fails before touching Telegram. The validated tag is stored trimmed, so the answer→--reply-to gate (a literal compare) can't be bypassed by a padded value like --tag " answer ".
• TAG_ALIASES (Cyrillic) removed; resolveTag uses a canonical Set and stays total for the renderer. The now-unreachable unknown-tag stderr warning + its resolveTag import were dropped from the entrypoint.
• Docs updated to lowercase-english only: tg --help / --format-help, the install-skill SKILL.md + always-on blurb, AGENTS.md, and the prefix.ts tag contract. The skill badge table now matches the real TAG_PILL_FALLBACK (🔵 ANSWER, …).

2. Colorized --help

tg --help and tg --format-help were the only ecosystem CLI helps printed flat. New features/render/help-color.ts colorizes section headers (bold cyan) and option names (green) to match review/rig/draw — dependency-free ANSI, the same \033[CODEm scheme as rig's _c. Auto-disabled when stdout/stderr is not a TTY or NO_COLOR is set, so piped/redirected help stays plain. Only color codes are injected: strip the escapes and the text is byte-identical.

Tests

bun test — 1014 pass / 0 fail. New coverage:

• validateTag accepts the four lowercase tags; rejects uppercase / mixed-case / Cyrillic / unknown with the 3-part error (unit + parseArgs + subprocess, asserting non-zero exit).
• the padded-answer-without---reply-to gate (regression guard).
• colorizeHelp headers + option flags get the right escapes; piped --help is plain (no \x1b[); strip-ANSI == input.
• install-skill badge table matches TAG_PILL_FALLBACK.
  Existing tag/render/reply-gate/install-skill tests migrated off Cyrillic + uppercase inputs.

Version bumped to 1.12.0 with a CHANGELOG entry.

Reviewed twice via review --staged (codex / opus / glm). Every finding addressed, including a real trim-vs-answer-gate regression the second pass caught (stored untrimmed tag bypassed the literal answer gate — now stored trimmed + tested).

Do not merge — CTO verifies.

🤖 Generated with Claude Code

[alex-mextner]

CI status — green except the pre-existing CodeQL legacy findings

All gates pass except CodeQL Self-Gate (javascript-typescript), which fails on 8 PRE-EXISTING findings in untouched code — not a regression from this PR:

rule	file (NOT touched here)
js/incomplete-multi-character-sanitization	features/auto-attach/transmitter.ts:51, features/render/rich.ts:244
js/double-escaping	features/auto-attach/transmitter.ts:52, tests/table.test.ts:124
js/incomplete-sanitization	features/render/rich.ts:87
js/file-system-race	features/install-skill/install.ts:218,248, features/transport/telegram.ts:67

These are the same 8 legacy findings tracked at #31 ("fix-or-justify the 8 CodeQL TS findings") — same situation as #25, which merged with this exact justification (the self-gate firing on legacy features/**, not a regression).

This PR's own new code introduced one CodeQL finding (js/incomplete-sanitization in tests/help-color.test.ts — a single-occurrence String.replace) — already fixed in commit 2bb4755, which dropped the gated count from 10 → 8. So everything this PR adds is clean; the residual 8 are #31's.

test gate is green (1014 tests pass).