docs: 8 spec improvements from building a PHP/Laravel implementation

[datashaman]

Description

Spec improvements discovered while building datashaman/attractor-app, a PHP/Laravel implementation of the Attractor spec. All changes are additive — implementation notes, missing mappings, and guidance for edge cases encountered in practice.

Changes across all three spec files:

unified-llm-spec.md:

1. Complete Gemini finish reason mapping — Added 6 missing values (BLOCKLIST, PROHIBITED_CONTENT, SPII, MALFORMED_FUNCTION_CALL, LANGUAGE, OTHER, FINISH_REASON_UNSPECIFIED) plus a note about 10+ newer image/advanced values. The existing table only listed 5 of 20+ possible Gemini values.
2. ContentFilterError clarification — Noted that this is a successful HTTP response (200) with blocked content, not an HTTP error. The raw provider reason (e.g., SAFETY vs SPII vs BLOCKLIST) matters for deciding whether to retry with a modified prompt or abort.

coding-agent-loop-spec.md:
3. Shell command escaping warning — Added explicit guidance that implementations must not escape shell metacharacters in the command string. In practice, escapeshellcmd() (PHP) and similar functions break pipes, redirects, subshells, and stderr redirection that LLMs routinely generate.

attractor-spec.md:
4. Async pipeline execution (new Section 3.9) — Guidance on job queue dispatch, progress reporting, timeouts, and failure recording. In practice, pipeline runs exceed HTTP timeouts within seconds.
5. DOT shape rendering note — Mdiamond/Msquare produce visual artifacts in browser-based Graphviz WASM. Noted that rendering layers may substitute shapes while preserving originals for handler resolution.
6. working_directory graph-level attribute — Every pipeline run needs a target working directory that flows to all node handlers. Added to the graph attributes table.
7. Observability decoupling — Implementation note that event emission should live in the orchestrator, not handlers. Handlers that call logging directly break in unit tests without a framework container.
8. Token aggregation across providers — When pipelines use different providers per node, token counts need per-provider subtotals since tokenizers differ and cost estimation requires provider-specific rates.

Breaking Changes

None — all changes are additive notes, table entries, and a new subsection.

[datashaman]

RE: 5. I see the geometric shapes play a part in the protocol.