chore: move measurements out of comments

The throughput and benchmark figures in the sink, chunk-size, hash-backend
and bulk-read comments restated what their commits already record, and a
number in a comment goes stale the moment the code around it improves. The
comments keep the constraint that still binds — what is slow, what must
align, what runs per byte — and the commits keep the numbers.

Raised by review (Greptile) on the PR.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>

apps/mobile/android/app/src/main/java/org/triliumnotes/trilium/TriliumFileSink.java

@@ -22,7 +22,7 @@ import org.json.JSONObject;
  *
  * <p>The Capacitor plugin bridge carries binary as base64 inside a JSON string, and the native
  * side re-parses that whole string per call ({@code MessageHandler} → {@code org.json}), which
- * caps file writes at roughly 13 MB/s. This listener receives raw {@code ArrayBuffer} messages
+ * is far too slow for large files. This listener receives raw {@code ArrayBuffer} messages
  * instead — no base64, no JSON around the payload — and appends them to a file on an IO thread.
  *
  * <p>Protocol, driven by {@code capacitor_download.ts}: a JSON string {@code {"type":"open",

packages/trilium-backup-container/src/backend-web.ts

@@ -9,9 +9,9 @@ import { KEY_BYTES, TAG_BYTES } from "./format.js";
 /**
  * The browser backend: AES-256-GCM and randomness from WebCrypto, gzip from the platform's
  * `CompressionStream`, and the two primitives WebCrypto does not offer from elsewhere — scrypt
- * from `@noble/hashes`, incremental SHA-256 from `hash-wasm`. The hash is WASM rather than
- * `@noble`'s pure JS because it runs over every payload byte and paced a phone's backup at about
- * 32 MB/s; scrypt stays JS because it runs once per container. Runs under Node just as well,
+ * from `@noble/hashes`, incremental SHA-256 from `hash-wasm`. The hash is WASM rather than pure
+ * JS because it runs over every payload byte, where the JS penalty on a phone paces the whole
+ * backup; scrypt stays JS because it runs once per container. Runs under Node just as well,
  * which is how the cross-backend tests exercise it.
  *
  * WebCrypto's `subtle` interface only exists in secure contexts (HTTPS, localhost, workers of

packages/trilium-core/src/services/sql/sql.ts

@@ -242,12 +242,10 @@ export class SqlService {
 
     /**
      * Gets every row of a whole-table read as a positional array, the way {@link getRawRows} does,
-     * but crossing between JavaScript and the database once instead of once per column.
-     *
-     * `@sqlite.org/sqlite-wasm` charges roughly 7µs per column a row exposes, so the Becca
-     * load spends most of its time marshalling rather than querying. `json_group_array()`
-     * hands the result set over as one value, which `JSON.parse()` turns into rows about ten
-     * times faster; better-sqlite3 is marginally faster this way too.
+     * but crossing between JavaScript and the database once instead of once per column: under
+     * `@sqlite.org/sqlite-wasm` every column read is its own wasm round trip, so a wide
+     * whole-table read spends most of its time marshalling. `json_group_array()` hands the
+     * result set over as one value for `JSON.parse()`.
      *
      * Going through JSON puts two limits on `columns` that {@link getRawRows} does not have:
      * `json_array()` rejects a BLOB column, and an integer above `Number.MAX_SAFE_INTEGER`