fix: fingerprint minidump crashes by crash site for Error Tracking

[Fiona2016]

Summary

• Minidump crash events now carry error.fingerprint, computed as {exceptionType}|{module basename}|{normalized module offset} from the first non-system frame of the crashed thread (e.g. SIGSEGV|MyApp|0x12ab3c).
• Without it, the backend groups every crash of one exception type into a single issue: all minidump crashes share the constant message Application crashed, so both fingerprint grouping and similarity grouping collapse distinct crash sites together.
• No backend change needed: the intake already stores error.fingerprint verbatim, grouping prefers an event-provided fingerprint, and similarity grouping is skipped when one is present (fc-rum logic/issue/fingerprint.go, group.go).

Details

• The offset, not the instruction address, identifies the crash site: ASLR rebases modules on every launch, while an offset is stable across runs of the same build. Offsets drift between builds, so a new app version opens fresh issues — the same trade-off as Android NDK top-frame grouping.
• Offsets are normalized (lowercase, leading zeros stripped, BigInt-exact) so equivalent spellings group together.
• Dumps without an identified crashed thread carry no fingerprint (and no stack); the backend's frames==0 gate already keeps them out of issue grouping.
• Assembly/transport needed no change: main-process events deep-merge all error fields and the batch transport serializes the event verbatim; only the RawRumError inline type needed the field added.
• CHANGELOG gets a [0.2.1] section, so the release workflow's notes extraction has something to find.

Review

Code-review pass performed against origin/publish before opening: no blocking findings.

Verification

• yarn vitest run src/domain/rum/error/CrashCollection.spec.ts — 31/31 pass (7 new: frame selection, system-frame skip, all-system fallback, no-fingerprint cases, offset-spelling equivalence)
• yarn test:unit — 696/696 pass
• yarn typecheck — clean
• yarn format — clean on changed files

Notes / Risks

• Repo-wide yarn lint has 9 pre-existing errors in files untouched by this PR (OperationCollection.ts, playground/*, e2e/*); CI does not gate on lint.
• Release coordination: this should ship as SDK 0.2.1 before fc-rum PR 👷 Update all non-major dependencies (minor/patch) DataDog/electron-sdk#162 (Electron crashes → Error Tracking) is deployed. Until 👷 Update all non-major dependencies (minor/patch) DataDog/electron-sdk#162 deploys, minidump crashes are dropped by the frames==0 gate and produce no issues, so there is no live over-merging to fix urgently.

[Fiona2016]

Reviewed — the approach and the implementation both hold up. Pushed one commit that corrects a comment; no behaviour change.

Checked and correct

• The site is keyed on module_offset, not an instruction or base address. Measured across three runs of the same crash: the module base moved every launch (ASLR) while the offset stayed put, so an address-based key would have opened a new issue on every crash.
• exceptionType is part of the key, so two different faults at one site stay apart.
• isSystemModule is given frame.module, which is the module's full path — formatFrameStack matches it against module.code_file, so the system-path checks actually fire. Test fixtures use real paths (/usr/lib/libSystem.B.dylib, /System/Library/...), so the skip-system-frame branch is genuinely exercised rather than passing by accident.
• Offset normalization is a good catch I had missed: equivalent spellings of one offset would otherwise split into separate issues, and BigInt keeps it exact past 2^53.
• Spreading the field rather than assigning it avoids an explicit undefined becoming an own property — the intake treats a present fingerprint as authoritative.

Verified end to end, since a fingerprint the backend never reads would be inert: the intake does store error.fingerprint on the error row, truncating at 255, and the column is VARCHAR(255). The longest realistic key here (EXC_BAD_ACCESS / KERN_INVALID_ADDRESS|Electron Framework|0x29e4bb4, ~66 chars) is well inside that.

The one fix: the note said the backend's zero-frame gate keeps a site-less crash out of issue grouping. That stopped being true earlier today — such a crash now reaches Error Tracking and groups by exception type and message. The code was already doing the right thing; only the stated consequence was stale.

Tests: 31 pass, typecheck clean.