refactor(refs): condense frontend-agent references to local knowledge (W-PR3)

[notque]

Track W, PR3 of 7. Frontend domain. Merges and condenses per the mandate — every agent keeps its references/ dir and at least one file. No agent deleted.

Measured

Agent	Before	After	Cut
typescript-frontend-engineer	17,247 w / 10 files	2,301 w / 3 files	87%
ui-design-engineer	9,628 w / 9 files	2,539 w / 3 files	74%
typescript-debugging-engineer	1,690 w / 1 file	unchanged	—
Total	28,565 w / 20 files	6,530 w / 7 files	77%

typescript-debugging-engineer was left alone deliberately: already a single file, already at budget. Its stale cross-refs into deleted typescript-frontend files were repointed.

Cut

React/TS tutorial material (hooks, discriminated unions, satisfies, template literal types), the composition-pattern catalog, full component-library implementations (Button/Modal/Tabs/Card/Table/Badge), WCAG 2.1 restatement, and default-Tailwind token boilerplate. The composition file was 100% reconstructible and went entirely.

Preserved

Both audits found no host incidents and no repo names in these two corpora — the asset here is operator taste plus version pins, and all of it survives:

• React View Transitions: every non-guessable rule (layout-vs-page placement, key semantics, flushSync skipping animations, persistent-element snapshot capture), both CSS workarounds (text-morph raster artifacts, backdrop-filter flash), the duration budget, and the full 12-row troubleshooting table.
• Next.js CVEs: CVE-2025-29927 middleware bypass (fixed 15.2.3), CVE-2025-55182 Server Action closure params, GHSA-rvpw-p7vw-wj3m image-optimizer SSRF, RSC prop leakage — each with its detection command.
• ai-slop-detection.md untouched. It is the operator taste, including every ripgrep detection command and the Japanese-indigo palette example.
• TS hard-gate table and its exceptions, anti-rationalization table, verification STOP blocks, death-loop retry limits, blocker criteria.
• A new consolidated non-obvious failure-mode table: inline-component identity churn, module-level state bleeding across concurrent renders, React.cache Object.is argument equality, RSC reference-identity dedup breakage, Strict Mode remount, useEffectEvent unstable identity, localStorage throwing in private browsing.
• 2-to-3 motion rule, 6-state matrix, transition timing bounds, 5-second test, whileHover-without-focus accessibility trap.

Defects fixed

• typescript-frontend-engineer.md declared not_for: twice — duplicate YAML key; the second silently won. Removed.
• Inter conflict resolved for the operator. design-tokens.md defaulted --font-sans to Inter (and implementation-patterns.md did the same) while ai-slop-detection.md names Inter as the tell of an absent typographic decision — the corpus failed its own ripgrep check. Both token files are gone; the surviving rule lives in ai-slop-detection.md, and design-conventions.md now states explicitly that a token set is not seeded with a default typeface.
• ui-design-engineer.md listed typescript-frontend-engineer — an agent — under "Companion Skills (invoke via Skill tool)". That call fails. Moved to a separate Companion Agents table dispatched via the Agent tool.
• golang-general-engineer-compact.md listed go-patterns twice (carried from PR2 review).

Gates (local, ruff 0.15.12 matching CI)

ruff check pass - ruff format --check 583 files pass - --check-do-framing pass - --check-placeholders pass - validate_reference_loading_tables.py exit 0 - pytest scripts/tests/ -q 4707 passed, 1 skipped, 128 xfailed.

--all structure issues: 5 agents / 10 issues, unchanged from post-PR2 baseline and 0 missing files (the two dangling cross-refs this PR initially introduced are fixed). All residual flags are NO_CODE_EXAMPLES advisories on agents this PR does not touch.

Pruned 3 now-deleted entries from _KNOWN_OVERSIZED (xfails 139 -> 128) and refreshed 2 reference-loading cases that named deleted files, per the same-PR rule.