[lars-1987]

Summary

Crash reasons now say where the crash happened.

 processFile
+  step = "read"
   file.arrayBuffer()
+  step = "strip"
   entry.strip(bytes, options, at)     # a processor may call at("...")
+  step = "assemble"
   assemble()
 catch
-  reason = `crash: ${name}`
+  reason = `crash: ${name} at ${step}`

stripPdf marks info:<Key> for each Info field, then xmp, custom and save. Other formats report read, strip or assemble until they mark their own.

Steps are fixed strings from our code, so the reason still never carries text from the file.

Evidence

• Before: a PDF whose CreationDate is Monday, 5 January gives crash: Error. That is the same reason as the production failures on 6 and 31 Oct.
  After: the same file gives crash: Error at info:CreationDate.

pdf with an unparsable CreationDate  -> "crash: Error at info:CreationDate"
pdf with text in ModDate             -> "crash: Error at info:ModDate", no file text in the reason
failed full read                     -> "crash: NotFoundError at read"
heic with a bad iloc count           -> "crash: RangeError at strip"

305 tests pass, typecheck and lint clean.

Merge Danger

Door: two-way

Blast Radius: telemetry

No user-facing change. The crash reason string changes, so queries matching exactly crash: Error need ILIKE 'crash%'. The malformed date crash itself is not fixed here.

[lars-1987]

Looks good