docs: add CLAUDE.md with repo-specific contributor and agent guidance

[DsThakurRawat]

Closes #327.

Adds a CLAUDE.md at the repo root. There was no CLAUDE.md, AGENTS.md or .claude/ anywhere before this, so anyone opening the repo with an AI assistant started cold, and a lot of contributors do start that way now.

It isn't a rewrite of the README or CONTRIBUTING. Those cover what the tool does and how to open a PR. This covers the things that are expensive to learn by hand.

What's in it

• What the package actually is, and the point that the evidence layer is the product rather than scaffolding, so *_cert / *_receipt / *_proof / verify* files don't get treated as dead weight.
• The install and test commands that match what CI does.
• Layout, with the sizes of the big subsystems, and a clear note that src/reflex/ is a compat shim where no logic should ever go.
• A "things that will bite you" section, which is the part with the most value:
  • Receipt-gated tests skip silently because receipt_test_support.py points at ../reflex_context/, which isn't in this repo. A green run is therefore not proof of parity ([C2] Receipt-gated parity/latency tests always skip in CI (green proves no parity) #292).
  • The quality ratchet is protected, fails closed, and a PR can't edit the workflows or the ruff/mypy sections of pyproject.toml to get around it.
  • The numpy and transformers pins each encode a real production failure in their comments, so they look removable but aren't.
  • uv.lock is committed but CI resolves floating deps with pip ([C4] uv.lock committed but unused in CI; pip resolves floating deps #294).
  • Tests are Linux only ([C8] Full test suite is Linux-only; macOS only smoke-tests doctor #298).
• Conventions: the stderr rule for CLI errors, the pytest markers, test-per-fix, and the two known inconsistencies (GOALS.yaml staleness [C6] GOALS.yaml ~3mo stale; infra resources still named reflex-* #296, license mismatch [C7] Packaging license mismatch: BUSL-1.1 declared, Apache classifier #297) flagged so nobody copies them forward.

Verification

Every factual claim was checked against the tree at 8bcfc68 rather than written from memory. Specifically confirmed: the ../reflex_context/ default in receipt_test_support.py:13, the ruff==0.15.10 and mypy==1.20.1 pins in the workflows, the mypy override list for tether.exporters.monolithic, the err_console / console split in cli.py, the file counts (239 test files, 143 scripts, 396 source files), and the current open state of every issue referenced.

Risk

Documentation only. One new file, no code and no CI changes, so nothing here can affect a build.

[DsThakurRawat]

@rylinjames flagging this one for you since it's mostly a record of decisions you made, and you're the right person to say whether I've written them down accurately.

Context: the repo has no CLAUDE.md or AGENTS.md, so contributors opening it with an AI assistant get no repo-specific context and rediscover the same traps by hand. #327 has the reasoning, this PR is the file.

Three things worth your eyes specifically:

• I've written down that a green CI run is not evidence of parity, because the receipt-gated tests skip silently against a ../reflex_context/ path that isn't in this repo ([C2] Receipt-gated parity/latency tests always skip in CI (green proves no parity) #292). That's an accurate description of the current behaviour, but it is a fairly blunt thing to state in a root-level doc. Happy to soften the wording if you'd rather it read differently, though I do think contributors need to know it.
• The quality ratchet section describes it as fails-closed and not editable from an ordinary PR. You built that over ci: add protected quality debt ratchet #316 and the follow-ups, so please check I've described the protected policy update path correctly.
• I've flagged the two open inconsistencies, GOALS.yaml being stale ([C6] GOALS.yaml ~3mo stale; infra resources still named reflex-* #296) and the BUSL vs Apache classifier mismatch ([C7] Packaging license mismatch: BUSL-1.1 declared, Apache classifier #297), mainly so nobody copies them forward. Tell me if you'd rather those stay out of a root doc until they're resolved.

Everything in the file was verified against the tree at 8bcfc68 rather than written from memory, so if something reads wrong it's likely my interpretation rather than a stale fact. Nothing here touches code or CI.