fix: the emergency 'Default' alphabet was selectable and dead (#88)

[willwade]

Live-reported bug

"as I drive it I dont see any text go into the output text writing area" — Dasher-GTK, tonight

Root cause, isolated by an engine-level repro on the reporter's actual profile: their settings held AlphabetID="Default" — the bare a–z "last ditch" fallback alphabet that CAlphIO's constructor registered unconditionally. That alphabet is structurally dead: with it selected, the engine renders, reports no error, and commits no output however long you drive (no space, no newline, no groups). Because it was registered like any real alphabet, it appeared in pickers' permitted values and round-tripped through settings.

The editor-contract work was exonerated by the repro: with a fresh profile the same binary drives text into the pane normally.

Fix

1. CAlphIO no longer eagerly registers "Default" — it disappears from pickers/permitted values. GetInfo falls back to the preferred real alphabet; the emergency alphabet is fabricated (once, lazily) only when nothing at all loaded — the case it was written for.
2. Realize() heals SP_ALPHABET_ID == "Default" to the loaded default, after the lazy alphabet-load block — pickers and the next settings save hold the healed id. Old settings self-heal on first launch after upgrade.
3. HandleParameterChange(BP_CONTROL_MODE): null guard. Setting control mode pre-realize was a latent null-deref (no NCManager yet); exposed by the create-flow shift, fixed properly — the NCManager ctor reads the flag, so pre-realize sets still apply at realize.

Two collateral traps dodged (documented in the investigation)

• Healing inside ChangeAlphabet broke the on-demand alphabet loader and set_string_parameter's stored-value contract (the "English, lower case" comma-normalisation wrinkle: name index vs parsed-name mismatch — worth its own issue).
• The guard-less control-mode case segfaulted only because the retired registration shifted when the NCManager first exists.

Tests

New regression case in dasher_alphabet_xml_tests: settings naming Default heal to 'English with limited punctuation' and the canonical drive produces output. Full matrix 43/43; clang-format clean.

Frontend follow-ups

• Dasher-GTK carries the live repro as an engine-contract regression test + submodule bump (PR follows).
• Reporter's immediate workaround (pre-merge): switch the alphabet picker off "Default" — e.g. "English with limited punctuation".

DCO signed.

Confidence Score: 5/5

The PR appears safe to merge; no actionable new defect or outstanding blocking finding remains.

Summary

• Lazily creates the emergency alphabet only when no real alphabet can be loaded.
• Resolves indexed alphabet files by their selected exact path.
• Prevents multiple C API contexts from reading each other’s data directories.
• Restores accumulated user training at startup without double-loading single-directory configurations.
• Guards control-mode changes made before realization.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Realize context] --> B[Scan this context's alphabet index]
    B --> C[Load configured alphabet by exact indexed path]
    C --> D{Stored ID is Default?}
    D -- No --> G[Create model and node manager]
    D -- Yes --> E{Preferred alphabet loaded?}
    E -- Yes --> F[Persist preferred real ID]
    E -- No --> H[Load real alphabets from this context's data directory]
    H --> I{Any real alphabet loaded?}
    I -- Yes --> J[Persist first loaded real ID]
    I -- No --> K[Retain last-ditch emergency behavior]
    F --> G
    J --> G
    K --> G
    G --> L[Load bundled training corpus]
    L --> M{User and data directories differ?}
    M -- Yes --> N[Load accumulated user training]
    M -- No --> O[Skip duplicate scan]

Loading

_{Reviews (5) · Last reviewed commit: "fix: on-demand load honours the indexed ..."}