Add selectable dictionary sources and fix unreliable word lookups

[hassanraza464]

Hi, thanks for the add-on. While using it I kept hitting HTTP Error 502: Bad Gateway on word lookups, and the digging turned up a second, quieter problem. This fixes both, then adds a way to avoid depending on a single dictionary.

Happy to split this into smaller PRs, drop the feature half, or adjust anything, just say.

The 502s

api.dictionaryapi.dev returns 200/404/502 nondeterministically per request. Their Cloudflare layer caches with max-age=14400, so a failure is served as a cache HIT for hours and looks deterministic; forcing a MISS with a cache buster showed the origin itself is unstable. Reproducible from unrelated networks, so it is not local to me.

The quieter problem

is, was and be return 404 from a fresh origin request, with the API's own {"title":"No Definitions Found",...} body, byte identical to what a nonsense word returns. The add-on cannot distinguish "no entry" from "not a word", so it was speaking "unable to find definition for word" for is. That behaviour survives the outage, since it is a property of the data rather than the downtime.

Changes

Reliability:

• Retry 500/502/503/504 and connection errors up to three times with a short backoff. 404 still fails immediately so unknown words stay responsive. Exhausted 5xx speaks a message rather than the raw urllib error.
• Verify a 404 against Wiktionary before reporting a word as unknown. Only if both sources agree is "unable to find definition for word" spoken; a word that exists but has no entry gives "definition unavailable". The check requires an en key, so pages existing only for other languages (e.g. hunde) are not mistaken for English words.

Word selection:

• Strip punctuation from the ends of a word before looking it up, so a word at the end of a sentence or inside quotes or brackets resolves. Internal apostrophes and hyphens are preserved (let's, so-so), the unstripped form is tried as a fallback for 'tis and e.g., and text that is only punctuation no longer triggers a lookup. Type detection still runs on the original text, so the IP, URL and ISBN paths are untouched.

Sources:

• A small provider abstraction returning normalised entries, so the formatting and audio code no longer cares where a definition came from. This is deliberately the shape dictionaryapi.dev already returns, so format_word_meaning and the audio dialog are unchanged.
• Wiktionary as a second selectable source, with HTML stripping, filtering to English entries, and flattening of its nested senses. Its parent senses embed sub-senses as nested <ol>, which duplicates every child inside its parent, and a parent's parsedExamples aggregates all of its children's (13 for be item 1, 31 for item 8), so examples are allocated per part of speech rather than taking the first available.
• A settings panel to pick the dictionary, plus an optional fallback to the other source. Defaults leave current behaviour untouched, so nobody's experience changes unless they open the panel.
• Pronunciation fallback to Wikimedia Commons when a recording will not play, which also came up during this outage since their media host was 502ing every pronunciation URL. It uses Wikimedia's mp3 transcodes, so the existing MCI player handles them and no new audio backend is needed. A failed playback now reports "pronunciation unavailable" rather than only beeping.
• Wikimedia responses are requested gzipped, taking a large entry from 69KB to about 9KB.

Also syncs manifest.ini with the version already set in buildVars.py, adds the two new modules to pythonSources so their strings are extracted, documents the settings in the readme, and fills in the new strings for es/fr/ru. I am not a native speaker of those, so please treat the translations as a starting point, or drop them if you would rather they went through the translators. No version bump, as that is your call.

Testing

Manually against NVDA on Windows 11, with both sources, the fallback on and off, and during the live outage. Verified the previously broken words (is, be, was) now return definitions from Wiktionary, that punctuation-adjacent selections resolve, that let's and so-so are unaffected, and that Commons audio substitutes when the dictionary's mp3 fails. Two pre-existing fuzzy entries, "Definition for {0}" and "Play {0}", are left as they were.

[hassanraza464]

Pushed two fixes found while testing:

• json.loads was outside the try in fetch_word_entry, so a 200 response with an empty body (which the API does while it is unwell) killed the lookup thread with a JSONDecodeError and the user was told nothing at all. Parsing is now inside the try, and an unusable body reports like any other failure. This also covers a valid but empty list, which raised IndexError on the [0].
• Looking elsewhere for a pronunciation ran between the beep and ui.message, so a lookup paused for roughly half a second before speaking whenever the fallback was enabled and the dictionary had no recording for that word. The definition is spoken first now and the recording fetched afterwards.

The second leaves a short window where a definition has been spoken but its audio has not arrived, so opening the dialog inside it shows no play buttons. Reaching it needs a double press faster than the lookup, and a second attempt has them. Happy to close that properly if you would rather, though every option I looked at trades it for either a delayed dialog or one that grows buttons while being read.