Persistence here has two rules and neither was tested. Losing the layout must
never lose the session, so every failure falls back to the first-run workspace
and carries on; and the fallback must be visible, so each path returns a
status string the footer shows. Nine different strings, one of which is the
only notice a user gets that the arrangement they spent a morning on has been
dropped, and not one of them was asserted anywhere. The way to find out that
a corrupt snapshot reports "invalid layout ignored" was to corrupt one.
Ten tests now cover the branches a real machine reaches: nowhere to write, a
data directory that cannot be created, a database SQLite refuses to open, a
fresh database that is ready rather than restored, a round trip that restores
a panel created before the save, and both ways a snapshot can be unusable --
malformed JSON and well-formed JSON describing a workspace with no attached
panel, since parsing is not validation and only the second is easy to write by
accident.
workspace_database_path now reads through EnvSource rather than std::env, for
exactly the reason that trait was introduced in lumbridge-settings: the
workspace forbids unsafe, set_var is unsafe in Rust 2024, and a precedence
rule that cannot be exercised without mutating the process running the test is
a precedence rule that stays untested. The one behavioural consequence is that
a non-UTF-8 value in LUMBRIDGE_SPIKE_DB, XDG_DATA_HOME or HOME is now treated
as unset rather than used as a path; that is what every other setting in the
shell already does with such a value.
persist_panels stays a method, reduced to the one thing the shell owns: a
workspace with no store is not a save failure. It is memory-only, the footer
has said so since startup, and replacing that standing message with an error
every time a pane moved would say less, not more.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SPYebLiN2w4TqnHUYGdECq
