Separate machine inference from embodied hearing

[emeisazam]

Summary:

• reserve heard for separately attributable human listening passes
• normalize model transcripts, captions, events, and host observations as inferred while retaining DSP as measured
• demote attempted machine heard claims at the gateway with the reason preserved
• align immutable component checkouts and bump Oída to 0.9.2
• broaden environment-file ignore coverage

Validation:

• local checks: 332 passed plus 13 subtests
• Ruff, compile, distributions, isolated wheel smoke, JavaScript syntax, release smoke, and dependency audit passed
• macOS Swift build passed
• staged diff secret scan found no findings