Keep the prompts that started the work, below the model rather than around it

Situation

The prompt that started the work is the one thing nobody keeps. It lives in a provider transcript that the next session cannot read, so six weeks later the repository holds the change and no record of what was asked for.

Scenario

setup: capture-installed
given:
  - 'Majordomus installed, with the provider hook written and the executable on PATH'
steps:
  - id: is-it-real
    run: ['capture', 'status']
    note: 'the state is read from the provider configuration and a payload driven through the shim, not from a flag somebody set'
    expect:
      exit: 0
      stdout_contains: ['claude-code', 'majordomus-capture']
  - id: one-prompt-one-record
    run: ['capture', 'prompt', '--provider', 'claude-code']
    stdin: claude-code-prompt.json
    note: 'what the provider sends on UserPromptSubmit; the record is the person half of the exchange'
    expect:
      exit: 0
      files_exist: ['.ai/local/prompts']
  - id: both-formats-or-neither
    run: ['capture', 'render']
    note: 'a prompt is kept twice under one stem: the record for a machine, the rendering for a person. render rebuilds a missing rendering from its record and never the other way round, because only that direction is possible'
    expect:
      exit: 0
      stdout_contains: ['rendering']
  - id: never-rejects-a-prompt
    run: ['capture', 'prompt', '--provider', 'claude-code']
    stdin: not-a-payload.txt
    note: 'in a hook, exit 2 rejects the person prompt, so a payload the tool cannot read is written to the failure log and the turn goes on'
    expect:
      exit: 0
      files_exist: ['.ai/local/prompts/.capture.log']
  - id: a-subcommand-that-is-not-one
    run: ['capture', 'nonsense']
    note: 'the command line still refuses what a person typed wrong: only the hook path is forbidden to exit 2'
    expect:
      exit: 2
      stdout_contains: ['unknown subcommand']
then:
  - 'the archive is written under the ignored half of the layer and never committed'
  - 'every prompt is present as both a JSON record and its Markdown rendering'
  - 'a prompt the tool could not read is a named failure, not a silence'

Outcome

The prompts are captured below the model, by the provider hook, into files the repository ignores and nothing loads into a context. Whether the wiring is real is answered by running a payload through it rather than by a claim, and a payload that cannot be read is reported instead of dropped.

Each prompt is kept twice under one stem, because the two readers are different. The .json record is what the provider sent — its own spans, still escaped, pretty printed one member per line — and it is what everything else is derived from. The .md beside it is that record rendered for a person: the fields as front matter and again as a table, then the prompt itself under ## PROMPT, decoded and fenced. An archive nobody opens is evidence of nothing, and a rendering nothing can be rebuilt from is not evidence at all, so the pair is enforced rather than left to habit. The asymmetry is what makes that safe: capture render rebuilds any rendering from its record, and no command can go the other way.