test(qwen): pin the orphaned-record dedup key shape

[ozymandiashh]

The smallest item from the review pass, and the one that says the most about the extraction's test coverage.

The pre-extraction decoder built dedup keys with template interpolation, so a record missing sessionId or uuid produced the literal qwen:undefined:<uuid>. Phase 8.1 changed the spelling to qwen::<uuid> by coalescing both fields to an empty string.

The intent was right — a missing identifier should contribute nothing rather than a fake value. But the change shipped with nothing behind it: no test, no fixture, no golden pinned either spelling, and the CLI parity golden only exercises fully-formed records, where the two are byte-identical. Nobody could tell the change from a defect. That is what this PR fixes — not the shape, the absence of a contract.

The test pins the orphan spellings, asserts the collapsed key still dedups identical records, and fails against the pre-migration spelling (verified red: qwen:undefined:a-orphan against the pinned qwen::a-orphan). The decode-site comment records the reasoning so nobody "restores" the old one.

Two caveats, worth knowing rather than discovering

• Records with an explicit null coalesce the same way as missing ones, so two previously-distinct keys now collapse into one.
• Keys persisted by pre-8.1 builds for orphaned records will not match the new spelling, so such a record can be counted once more across the upgrade. The qwen corpus is not in the frozen golden set, so no shipped fixture moves.

[ozymandiashh]

Converted to draft: this was opened with only a self-review by the model that wrote it, which is not the review bar this batch is held to. Independent review by two other reviewers is running now; I will mark it ready once both have passed and any findings are applied. Apologies for the noise.