Stop on a malformed config or bindings file

[844196]

Closes #59

Summary

A broken config file was indistinguishable from an absent one, so a typo silently changed how wk behaves. Deno.errors.NotFound is now the only silent fallback; anything else — a syntax error, a shape mismatch, EACCES, EISDIR — raises a ConfigError and exits 7.

Background

Condition	Before
syntax error in config.yaml	falls back to the defaults, exit 0
syntax error in bindings.yaml	treated as empty, every key undefined, exit 5
bindings.yaml holds a mapping or a scalar	no error — [].concat({foo:'bar'}) yields a one-element list and the bogus entry reaches the menu
config.yaml holds a scalar (42)	no error — mergeContext(42) passes

Two causes. .catch(() => fallback) in run.ts could not tell "the file is not there" from "the file is broken", and nothing validated the shape of what was parsed.

Changes

• src/errors.ts: add ConfigError, carrying the path and the reason.
• src/run.ts:
  • rewrite loadYaml: only NotFound falls back, everything else raises ConfigError.
  • an empty document (null, comments only, ---, ~) still falls back, keeping the behaviour (config, bindings): an empty config file crashes instead of falling back #58 landed.
  • the three files are read in a fixed order, one at a time — config.yaml, the global bindings.yaml, then the local wk.bindings.yaml — so the first broken file is the one reported and the rest are left unread. The previous Promise.all concurrency is dropped; it bought nothing, because tui.init()'s cursor-position query is a synchronous blocking read that stalled those loads anyway.
  • one line on stderr, with no wk: prefix since the widget adds its own, and $HOME abbreviated to ~:
    • <path>: <first line of the @std/yaml message> for a syntax error — it already carries at line N, column M; the excerpt and caret that follow are dropped so the message fits zle -M.
    • <path>: invalid format for a shape mismatch.
• src/widget.eta: comment only. Its *) arm already forwards an unknown exit code to zle -M, so no behavioural change.

Scope of the validation

A small hand-written check, so no new dependency:

• config.yaml must be a non-null, non-array object.
• both binding files must be arrays whose elements are non-null objects with a string key.

Field types (type, buffer, …) and nested bindings: are deliberately not checked.

Breaking change

Yes. A malformed config.yaml used to exit 0 and let wk keep running on the defaults; it now exits 7 and refuses to start. Input that used to succeed now fails, so this belongs in v2.0.0.

This PR alone does not ship v2.0.0. The field-type validation left out above is to be added first, and v2.0.0 released after that.

Verification

• mise run check passes.
• The full e2e suite passes (70 tests). e2e/tests/04_config.bats is now 19 tests: the two that pinned the silent fallbacks are updated to exit 7, and eight cases are added —
  • the local wk.bindings.yaml fails fast just like the global layer,
  • three shape mismatches (a mapping, a missing key, a scalar),
  • config is read before bindings, so the first broken file wins,
  • a directory in place of config.yaml,
  • a path under $HOME is reported with a tilde.

The new assertions match on the path prefix and at line N, column M rather than pinning @std/yaml's wording or the errno string verbatim, since e2e/helpers/common.bash states the suite may not reference Deno and doubles as an acceptance spec for any future reimplementation.