Troubleshooting guidance depends on injection logs that are disabled by default (v0.45.1)

README.md:292, in the "file an issue" checklist, asks reporters for:

What the plugin injected (or didn't) — enable debug logs with VERCEL_PLUGIN_LOG_LEVEL=debug

That cannot produce output on a default install, because the hooks that emit
those logs are not registered.

hooks/hooks.json registers four hooks:

SessionStart -> session-start-seen-skills.mjs
SessionStart -> session-start-profiler.mjs
SessionStart -> inject-claude-md.mjs
SessionEnd   -> session-end-cleanup.mjs

user-prompt-submit-skill-inject.mjs, pretooluse-skill-inject.mjs,
prompt-analysis.mjs, unified-ranker.mjs and lexical-index.mjs all
instantiate the logger but are absent from hooks.json, and no registered hook
imports or spawns them (checked both).

I understand this is intentional — README.md:36 says skills "are no longer
auto-injected on every tool call or every prompt by default" and the engine is
kept "for targeted or future opt-in workflows". The problem is that the rest of
the docs don't reflect it.

Measured on v0.45.1, running the registered hooks directly with
VERCEL_PLUGIN_LOG_LEVEL=trace:

session-start-seen-skills.mjs   stderr: 172 bytes   (one SessionStart line)
session-start-profiler.mjs      stderr: 0 bytes
inject-claude-md.mjs            stderr: 0 bytes     (no logger at all)

So trace — the loudest level — yields a single line about session state and
nothing about injection. A user following README:292 sees silence and cannot
distinguish "plugin decided not to inject" from "logging is broken" from
"plugin never ran".

Docs that read as currently-active but are not, by default:

• CLAUDE.md:52-59 — describes the SessionStart → PreToolUse → UserPromptSubmit
  injection flow in the present tense, including budgets and dedup
• CLAUDE.md:170-175 — documents VERCEL_PLUGIN_INJECTION_BUDGET (18000),
  VERCEL_PLUGIN_PROMPT_INJECTION_BUDGET (8000) and
  VERCEL_PLUGIN_TSX_EDIT_COUNT as live knobs
• vercel.md:963 — documents VERCEL_PLUGIN_HOOK_DEDUP=off, read only by
  prompt-analysis.mjs and user-prompt-submit-skill-inject.mjs, so it is
  currently inert

Suggested fix, in rough priority order:

1. Amend README:292 so the troubleshooting checklist doesn't depend on logs that
   are off by default — or point at whatever is observable (the SessionStart
   line, active-session.json).
2. Note in CLAUDE.md's Skill Injection Flow that steps 2 and 3 are not
   registered by default.
3. Document how to opt in, which README:36 implies is supported ("targeted or
   future opt-in workflows") but does not explain. If there is no supported
   opt-in yet, saying so explicitly would help — the env knobs currently read as
   though there is.

Not asking for the injection hooks to be re-enabled; the lightweight default
seems deliberate and reasonable. This is only about the docs matching it.

Environment: vercel plugin 0.45.1 (claude-plugins-official), macOS, Claude Code.

Start with README.md:292 and :36, CLAUDE.md:52-59 and :170-175, vercel.md:963, and hooks/hooks.json; compare the documented behavior with the registered hooks and the named injection entry points. Update the documentation to distinguish default behavior from supported opt-in behavior and make troubleshooting observable. Done means the docs no longer present inactive defaults as live and clearly explain any supported opt-in path.