Fix robustness bugs in gmd(): clean errors, flag coercion, year filtering, fetch hardening

[Yicheng-Zhao1999]

Summary

Fixes the batch of robustness issues reported for the Python global_macro_data package. Each fix was developed step-by-step with an independent review pass, and the whole set was re-verified against the live backend.

Test suite: 130 passed (99 baseline + 31 new regression tests), fixture-backed and deterministic.

Bugs fixed

#	Symptom (before)	Behavior (after)
1 (P0)	A listed version whose .dta is missing failed with a generic "raise an issue"	Clear GMDCommandError naming the version & path: The data file for version X could not be retrieved (distribute/GMD_X.dta)…
2 (P1)	country=840 (scalar int) leaked TypeError: 'int' object is not iterable	Clean GMDCommandError: country must be a string or a list of strings…; lists containing ints still accepted
3 (P1)	get_available_versions() leaked a raw RuntimeError when offline	Wrapped in GMDCommandError; cache/local fallbacks preserved
4 (P1)	raw="no" / iso="off" were truthy and activated; fast only matched "yes"	New _coerce_flag: accepts bools + a whitelist (yes/no/true/false/on/off/1/0); ambiguous values raise; fast now consistent
5 (P2)	start_year= leaked TypeError; no year filtering at all	start_year/end_year are real params doing inclusive row filtering; genuinely-unknown kwargs still raise TypeError
6 (P2)	Wrong setup.py author; unpinned deps; no User-Agent/retry; cache could be poisoned by a partial write	Correct author/email; pinned deps; _fetch_from sends a User-Agent and retries with backoff (no retry on 4xx); _atomic_to_stata writes the cache atomically

Extra: GitHub-mirror fallback for helper tables

Remote reads now try the S3 release bucket first and fall back to the GitHub mirror. Note (documented in-code): the GitHub mirror only hosts the helpers/* tables — it does not host the versioned datasets (distribute/GMD_<ver>.dta) or per-variable raw CSVs, so those remain S3-only and cannot fall back. Verified live: with S3 forced down, helpers/versions.csv still loads from GitHub while a dataset fetch fails cleanly.

Notes for maintainers

• The original Bug 1 S3 404s (2025_01/03/05/06/08) no longer reproduce on the live bucket — all nine versions return 200 and load end-to-end, so the missing vintages appear to have been re-uploaded already. The code-side improvement (clearer message) is still worth keeping. Suggest a release-pipeline pre-assertion that every version in versions.csv has a reachable distribute/GMD_<ver>.dta.
• _fetch_primary is now unused by production code (kept for test stability).

🤖 Generated with Claude Code