[Paritsingla7]

Summary

Phase 0+1 of the schema-diff idea (comparing Hyprland's API between two versions to eventually warn a config about upcoming breakage, not just validate against whichever one version it's pointed at today). Deliberately foundation only — no CLI surface — wanted the core comparison logic validated against real version transitions before deciding how to expose it.

• schemas/v0.55.0.json .. v0.56.2.json — one real schema per tagged Hyprland release (all 8 that exist), not synthetic fixtures. Cross-checked against schema.json already in this repo — identical.
• scripts/generate_schema_history.py — regenerates/extends the corpus. Confirmed generateLuaStubs.py is pure static analysis over C++ source text (no build needed), and that it must be run as each tag's own copy against that tag's own source, since the generator's own parsing logic has changed across versions. Verified end-to-end against the live Hyprland repo — regenerating a removed schema reproduced it byte-for-byte.
• hyprvalidate/schema/diff.py — diff_schemas(): classes/fields/aliases/config-keys added, removed, type-changed, plus a conservative "possible rename" heuristic (fires only on a clean 1:1 same-type match, never guesses past that).

Grounded in two real, dated Hyprland regressions

• hl.permission{}'s allow → mode (hyprwm/Hyprland#14400, broke one day after Lua config first shipped) — a clean 1:1 match the heuristic catches confidently.
• HL.Window's over_fullscreen → allowed_over_fullscreen (hyprwm/Hyprland#15367) — genuinely ambiguous by type alone (two new booleans were added, not one) and is deliberately left as a plain removal rather than guessed at. Asserting this negative is the point: a wrong confident guess is worse than an honest gap.

Every diff is phrased as "the schema changed," never "this will break" — the generated stub has itself lagged real Hyprland behavior more than once (the allow/mode field was wrong from the version it first shipped in), so a schema diff is evidence of a possible behavior change, not proof of one.

Test plan

• pytest tests/ — 203 passed (3 pre-existing/environmental entry-point failures unrelated, deselected)
• scripts/generate_schema_history.py run end-to-end against the live Hyprland repo (temporarily removed one schema, regenerated it, byte-identical to the original and to the repo's own schema.json)
• Tag-discovery logic verified to exclude ancient pre-rewrite tags (e.g. v0.6.0beta) and find exactly the 8 real Lua-era releases