chore: declutter repo, and fix the documented pandas API

[OmerBaddour]

Two things: a cleanup, and a public-facing docs bug the cleanup uncovered.

1. Decluttering

Path	Why removed
.claude/agents/, .claude/agent-memory/	The multi-agent layer was never useful, and its memory had rotted — it pointed at src/clair/auth/environments.py, which does not exist (real path: src/clair/environments/).
.claude/specs/pandas_trouve*.md	822 lines describing an abandoned design (see below).
docs/	PRD + architecture for clair docs, a shipped feature. Its logo was byte-identical to the site_docs copy.
specs/environments.md	Feature shipped; site_docs documents it.

Renamed site-docs → site_docs per the underscore convention (workflow, .gitignore, README updated). Markdown filenames keep hyphens — they are public URL slugs.

2. Memory, rearchitected

.claude/memory/ now holds only what documentation cannot: project_design_invariants.md (seven rules, each re-verified against code) and project_docs_are_the_source_of_truth.md. Routing-collision and environments notes dropped as duplication of guides/routing.md and concepts/environments.md.

CLAUDE.md gains a "What clair is" section, a package-level map of src/clair/, and a 3-step opening move for feature requests (grep site_docs → read MEMORY.md → open only the named files). One line per package, so it stays true as files move.

3. The pandas API was documented wrong

PandasTrouve does not exist. from clair import PandasTrouve raises ImportError — and README.md is the PyPI long description, so the first code block on the package page was broken.

The real API is a df_fn on the ordinary Trouve, with dependencies declared as parameter defaults:

def daily_event_counts(
    refined_events: pd.DataFrame = upstream,  # type: ignore
) -> pd.DataFrame: ...

trouve = Trouve(df_fn=daily_event_counts)

The docs were written from the design spec, whose two headline decisions — separate class and inputs dict — were both reversed during implementation. The result was self-contradictory docs:

Surface	Documented	Correct?
reference/trouve-api.md, concepts/trouve.md	df_fn	✅
guides/pandas-native.md (12×), index.md, README.md (5×)	PandasTrouve	❌

All now use df_fn. Three further claims corrected against real output from example_projects/example_4:

• compile writes a .py artifact, not a .json manifest
• clair dag renders roots down to dependents (the old example had the arrows backwards) and tags nodes [PANDAS]
• there is no clair[pandas] extra — pandas is a hard dependency

This also disproved a claim I had written into the new memory note ("CI publishes the docs, so they stay correct"). mkdocs does not execute examples. The note now states that the code is the final authority.

Checks

• uv run pytest tests/ — 516 passed (no code changed)
• uv run --group docs mkdocs build --strict — clean, no broken links
• clair compile and clair dag run against example_projects/example_4 to verify every documented claim
• No PandasTrouve reference remains in README.md or site_docs/

🤖 Generated with Claude Code