fix(core): name the file and key in config value type errors

[ElbertePlinio]

Closes #46

validate_keys in crates/core/src/config.rs now type-checks each known value in each config file before merging. It deserializes the value into the same type its Config field uses, so the accepted set is unchanged. A wrong type reports `<key>` in <file> must be <expected>, where the expected text is "a non-negative integer or null" for global and per-language limits, "a non-negative integer" for hook.max_blocks, and "a list of strings" for ignore, tests.patterns, and tests.exempt. The final deserialize of the merged config stays as a backstop. docs/spec.md states the rule, and the release draft gains a line for #46.

One behavior change goes beyond the message. Main type-checked only the merged config, so a bad value in the user config was silently dropped when the repo config replaced the same languages.<name> entry (QA 01 exits 0 on main). Each file is now checked on its own, the same way unknown keys already were, so that config now exits 2 and names the user file.

Requirement	Verified by	Scenario	Result	Evidence
A value type error names the config file and the full key path	automated test + CLI QA	string, float, boolean, negative, and out-of-range limits; per-language string limit in the user config; hook.max_blocks of -1 and null; ignore as a string or null; mixed tests.patterns; string tests.exempt; non-string tests.exempt entries [1] and ["lines",1]; bad value in the repo config with a valid user config	pass, exit 2 with key and file	string_global_limit_names_key_and_config_path, string_language_limit_names_full_path, float_limit_is_rejected, negative_max_blocks_names_key_and_config_path, string_ignore_is_rejected_as_non_list, non_string_test_pattern_is_rejected, string_test_exempt_is_rejected_as_non_list, non_string_test_exempt_entry_gets_list_message; QA 01 to 12, 17
Tests cover a wrong type in a global limit, a per-language limit, and hook.max_blocks	automated test	the tests above, asserting the whole message with the temp config path	pass	crates/core/src/config.rs
Valid configs and other errors unchanged	automated test + CLI QA	null global and per-language limits, zero limits, max_blocks 0, valid ignore, empty config, unknown limit key with a bad value, non-object per-language limits, tests.exempt of ["depth"]	byte-identical to main, including exit codes	null_limits_still_load and existing config tests; QA 13 to 16, 18 to 20
Tests, lint, format, coverage, PickCheck	automated	cargo test --workspace --locked --all-targets, cargo clippy ... -D warnings, cargo fmt --check, cargo run -- check crates, cargo llvm-cov --fail-under-lines 94 (94.92%), pickcheck check --base origin/main --fail-on new,worsened,unmatched (0 violations), on b48c1c5	pass	CI on this PR

QA: a CLI pass ran release builds from main (213fbbd) and this branch over a scratch Git repo in 20 scenarios, with a user config supplied through XDG_CONFIG_HOME, plus --changed and --base HEAD --format json with a string limit. The 13 wrong-type cases now name the key and file, and the 7 valid-config and other-error cases are byte-identical to main. After the review fix, the tests.exempt cases were rerun: [1] now gets the list message instead of unsupported tests.exempt key `<non-string>` , while ["depth"] keeps its old message. The evidence stays local and outside the repo.

The coverage floor stays at 94, which is 94.92% rounded down.

Review: independent xhigh reviews by Astra and Fable at d2c873b. Astra found nothing. Fable found no blockers. Its minor finding was that a non-string tests.exempt entry kept the older <non-string> message. Its nits were a needless clone in the type check and a test that bundled three scenarios. b48c1c5 fixes all three, and the Fable recheck confirmed it. A nit to unify backtick style with the older shape messages was declined, because the new messages follow the existing unknown-key style and restyling older messages is outside this issue.