fix(config): migrate deprecated keys in env vars and inline tables

[max-sixty]

The gap

#3152 routed the --config-set layer through the deprecation-migration layer, but a deprecated config key could still slip through unmigrated in two other places. Both are now closed with the same mechanism (migrate_content).

Env-var layer

migrate_env_overlay() runs the assembled WORKTRUNK_* overlay through the same migrate_content() config files and --config-set use, before the deep-merge. So WORKTRUNK__MERGE__NO_FF=true resolves to the deprecated key merge.no-ff and now takes effect as merge.ff=false, instead of falling through as an unknown field and being silently ignored. Migrating before the merge (not after) keeps the env value winning over a lower layer's canonical key, matching the layer precedence.

One behavior edge, now documented on the function and pinned by a test: canonicalization surfaces a type mismatch the deprecated name hid. WORKTRUNK__COMMIT_GENERATION__COMMAND=42 migrates to commit.generation.command = 42, which fails to deserialize (the field is a String), so the whole env layer is dropped with an attributed LoadError::Env warning and file config is preserved — the same contract a type-mismatched value in a canonical env var already follows. Previously the unknown key was silently ignored and other env vars survived.

Inline-table form

migrate_negated_bool_doc resolved its section with as_table_mut(), which returns None for an inline merge = { no-ff = true }, so the inline form was never migrated (and surfaced as a double unknown-field warning instead of a deprecation warning). It now branches on standard vs inline table, fixing config files, --config-set, and env at once — the inline shape is preserved (merge = { ff = false }).

remove_switch_picker_timeout_in had the same as_table_mut()-on-switch limitation, so it's given matching inline-switch handling and now every switch sub-key rule is inline-aware. With that, the DEPRECATION_RULES ordering comment is corrected: the [select] → [switch.picker] rule's inline-to-standard switch conversion is no longer relied on by any rule; only its timeout-ms move remains load-bearing.

Tests

• Unit: test_env_overlay_migrates_deprecated_key (full EnvVar → resolve → migrate path), test_migrate_negated_bool_inline_table.
• Integration: test_switch_env_var_migrates_deprecated_no_cd (cd-directive bug-catcher — reverting the fix re-adds the "Cannot change directory" line), test_list_config_env_deprecated_type_mismatch_drops_layer (the documented whole-layer-drop edge).
• Inline cases added to the test_warning_fires_iff_update_changes battery (negated-bool inline, inline-switch timeout-ms).

This was written by Claude Code on behalf of max

[worktrunk-bot]

Approved. One advisory check went red — affected tests (macos, advisory) — but it's an unrelated flake, so the approval stands.

The failure is a 180s timeout on integration_tests::shell_wrapper::unix_tests::test_source_flag_forwards_errors::case_1, a shell-wrapper PTY test. It's unrelated to this PR (pure TOML deprecation-migration logic, platform-independent): the same advisory job passed on linux and windows, the required test (macos) passed, and the cause is a non-deterministic timeout rather than a build/assertion error. A rerun of that leg should clear it.

The required test (windows) was still in progress when I stopped polling — unverified here, but required CI on linux/macos plus codecov/patch are green.