Feat: extend the retry NDJSON event with correlation fields, cliVersion and a completion event

[byapparov]

Context

#115 (backport of #110) added retry and execution-outcome telemetry: cliVersion on every event, retry_scheduled / retry_complete events, and an interrupted reason. #122 has since landed a retry NDJSON event {attempt, code, next}, backed by SessionRetry.Reason (no_finish_reason, rate_limited, overloaded, server_error, network, unknown) and documented in EVENTS.md. The two designs clash:

• they have different reason vocabularies (rate_limited vs rate_limit, server_error vs provider);
• merging both would emit three events per retry.

#115 is closed in favour of building on #122.

Goal

Keep #115's useful parts on top of the existing retry event:

• cliVersion on every event, or on invocation_start / session_start if that's enough;
• correlation fields on retry, e.g. the step or message id and the underlying error class;
• one completion signal for a retry sequence (recovered, or exhausted with its terminal error);
• the interrupted terminal reason, if it's still needed after fix: propagate provider finish errors to headless runs #113 and fix(session): retry steps that end without a finish reason (#121) #122.

Acceptance Criteria

• One retry vocabulary: SessionRetry.Reason stays the single source, and the EVENTS.md / SDK parity test from fix(session): retry steps that end without a finish reason (#121) #122 still passes.
• No duplicate retry events: one retry per attempt, plus at most one completion event.
• Additive under the EVENTS.md schema policy, and documented.
• Raw provider text never appears in these events (fix(session): retry steps that end without a finish reason (#121) #122's sanitisation stays in force).

References

• Supersedes feat: expose retry and execution outcome telemetry #115. Builds on fix(session): retry steps that end without a finish reason (#121) #122 (issue Fix: retry a step that ends without a finish reason, and never send Gemini a trailing model turn #121).