docs(tutorial): fix step 15 candidate-vs-baseline coherence

[placerda]

Why

Step 15 of the prompt-agent quickstart instructed the reader to paste
back a byte-identical copy of the baseline prompt as "the fix". That
made prompt_deploy stage report action: "reused" and keep
candidate_agent: travel-agent:2. When the reader then looked at the
Release Evidence card and saw Target: travel-agent:2, the natural
reaction was "I just shipped a fix — why is nothing new here?".

The card was correct (the Release Evidence Target mirrors what
agentops.yaml pins — the declarative baseline), but the tutorial
never made the candidate-vs-pin distinction explicit, so the story
read as "the fix didn't take".

What this PR changes

Docs-only. No code, no schema, no behavior.

• Mental model section. Adds a 4th bullet that names the split:
  PRs create and evaluate a new candidate Foundry version; the
  agentops.yaml agent: pin (and therefore the Release Evidence
  Target field) only moves when you explicitly promote a baseline.
• Journey table row "Fix and redeploy". Updates the description
  to match the new step 15: push an improved prompt that restores the
  guardrails and adds a small refinement, so a brand-new candidate
  version is created (instead of a reused no-op).
• Step 15 itself.
  • Replaces the byte-identical restore prompt with an improved
    prompt: keeps all four original structural rules, adds one narrow
    rule ("include currency code when known"). The change is small
    enough that similarity scores against the existing dataset should
    stay above thresholds, but real enough that
    prompt_deploy stage reports action: "created" and a new
    candidate_agent: travel-agent:N.
  • Rewrites the explanation block to walk the reader through the
    new foundry-agent.json, the candidate-vs-pin split, and where
    durable cross-environment identity lives (prompt_sha256,
    git_sha).
  • Adds a recording-note callout: if the reader already pushed a
    byte-identical restore in an earlier take, push one more commit
    with the improved prompt and the stage step will report
    action: "created" instead of reused.
• Success criteria checklist. Updates the "you restored the
  prompt" bullet to mention the new candidate version and the
  unchanged baseline pin, so the checklist matches step 15.

Why a narrow additive change for the new prompt

A purely-restorative "fix" can't show the
action: "created" + new-candidate-version part of the story — by
definition, byte-identical content reuses the existing version. We
need the fix to be a real content change. We also can't change the
prompt too much, or we risk dropping similarity scores below the
PR-gate thresholds and turning step 15 back into a failing run.

The chosen new rule (require currency code when monetary amounts are
mentioned) is narrow and additive, keeps the four structural
guardrails intact, and stays in scope for the dataset used by the
tutorial.

Release impact

None — this is a docs change. Will land on develop and ship with the
next release cut.

Test plan

• Inspected the rendered Markdown locally; numbered headings and
  table formatting still align.
• Grepped for stale "Restore"/"restored prompt" wording — only
  the success-criteria checklist line was affected and is
  updated.
• No code paths changed; no test suite run required.

Related context

• Surfaced by user feedback while recording the quickstart video: the
  Release Evidence Target: travel-agent:2 after the "fix" felt
  contradictory.
• Confirmed root cause by inspecting the foundry-agent.json
  artifact from
  https://github.com/placerda/agentops-prompt-quickstart/actions/runs/26714275682
  (action: "reused", candidate_agent: travel-agent:2,
  prompt_sha256 identical to baseline).