adf: no repo records whether vendor files are compiled or pointers, so compile --check and tidy disagree forever

[stackbilt-admin]

Split out from #296, which fixes the immediate data loss. This is the design question underneath it.

The problem

adf compile --write expands .ai/ into the vendor files. adf tidy folds the vendor files back into .ai/. They are exact inverses, and nothing records which one a repo has chosen.

#296 stops tidy from eating compiled output. It does not answer "which mode is this repo in?", and that question has other consequences.

Evidence it is not hypothetical: this repo

Charter's own vendor files are thin pointers:

CLAUDE.md        tracked=yes  pointer (18 lines)
AGENTS.md        tracked=yes  pointer (8 lines)
.cursorrules     tracked=yes  pointer (5 lines)
GEMINI.md        tracked=yes  pointer (8 lines)

And adf compile --target all --check reports all four stale, permanently:

[stale] CLAUDE.md — out of date with .ai/ source.
        Run: charter adf compile --target claude --write
[stale] AGENTS.md — out of date with .ai/ source.
...

That remediation line is wrong for this repo. Running it would replace the deliberate pointers with full compiled content — changing the repo's mode, not fixing drift. --check has no concept of pointer mode; it assumes compiled mode unconditionally and calls any pointer stale.

This is the same root cause #296 reports as a "related symptom", not a separate defect: the reporter saw permanent staleness because tidy had turned their files into pointers.

Why the documented path walks into it

README.md:39 tells every new user:

npx charter adf compile --target all --write

then hook install --pre-commit, then commit — at which point the hook runs tidy. The quickstart puts a fresh repo into the oscillating state. #296 stops that from corrupting .ai/, but the repo still ends up with --check permanently red and a remediation that would flip its mode.

Options

1. Record the mode in .charter/config.json — e.g. vendor: "compiled" | "pointer". tidy no-ops in compiled mode; compile --write warns in pointer mode; --check only checks compiled-mode repos, and in pointer mode verifies the pointer is intact instead. Explicit, greppable, survives hand-editing. Costs a config migration and a default for existing repos.
2. Detect per-file from the compile banner — the approach bug(adf tidy): ingests compiled vendor output back into core.adf, duplicating content and reverting deliberate deletions #296 takes. Zero migration, matches what .githooks/pre-commit:16 has done in the field. But a repo can sit mixed indefinitely, and stripping the banner silently flips a file's meaning.
3. Make them explicit inverses, last writer wins — most honest about intent, largest change, and does not help the pre-commit hook running tidy unattended.

Recommendation recorded during the #296 triage: 2 now (shipped in #296), 1 later. This issue tracks 1.

Also worth fixing when this is picked up

charter adf sync --write is a natural thing to reach for when --check says stale, and it is not the fix — it writes .ai/.adf.lock from declared SYNC: entries and reports Wrote empty lock file (no SYNC entries declared) when there are none. That message is accurate but lands as a dead end for someone chasing a stale compile --check. Worth a pointer in the output toward what staleness actually means once mode is a concept.

Acceptance

• A repo can state its vendor mode, or it is inferred consistently by every command that cares (compile --write, compile --check, tidy, the generated pre-commit hook, doctor).
• compile --check does not report a deliberate pointer as stale.
• No remediation line suggests a command that would silently change the repo's mode.