llm codec: translated bytes back to 1.5.5 (row-159 L3#5-#9, L2#1)

[MattJackson]

What. These are codec HIGHs from the row-159 plane-llm audit. Each puts back the bytes 1.5.5 sent, or fixes a 1.6.0 defect that has no 1.5.5 counterpart. None of them moves a billed figure. The one billing edge, Gemini audioDurationSeconds on the read side, is split into its own $ PR.

finding	change	1.5.5 source
L3#5	A translated stream into a Responses client stamped created_at: 0 (1970). It now stamps the host wall clock (busbar_contract::codec::wall_clock_now, the same reading the Chat writer uses).	v1.5.5 proto/openai_responses/writer.rs:479 (now_unix_secs())
L3#6	The same-dialect relay's stop-reason scan matched the first "finishReason" anywhere in the body, so a model-written functionCall.args object faulted a shared lane. StopKey now names a member at a container depth (Gemini candidate = 3, Cohere top level = 1), and the scanner walks the JSON lexically, inline and without heap use.	Q31 failure detection kept
L3#7	Gemini candidatesTokenCount was the output minus reasoning, with a new thoughtsTokenCount. It is the whole output again, with no new member.	v1.5.5 proto/gemini/writer.rs:962-964
L3#8	Streamed Gemini tool ids after text counted only a still-open text block (position 2 for the second call). They now count text_index.is_some(), matching the buffered ids.	v1.5.5 proto/gemini/reader.rs:895
L3#9 (writer)	A duration-billed transcription no longer writes an invented usageMetadata.audioDurationSeconds.	v1.5.5 handlers/gemini.rs:259-265
L2#1	A lane without native structured outputs drops response_format again, with a warn and a reported control. It no longer sends a synthetic busbar_response_format forced tool, which overwrote the caller's tool_choice and streamed back to the client as a tool call.	v1.5.5 proto/anthropic/writer.rs:343-347; CHANGELOG 1.6.0 "an omitted key sends what 1.5.5 sent"

RED-able tests.

• tests/proto/stream_translate_tests.rs::a_translated_stream_into_responses_carries_the_host_clock_created_at
• gemini/tests/tests.rs::{a_finish_reason_inside_tool_args_is_not_the_stop_reason, stop_key_scanner_reads_the_candidate_member_at_every_split, streamed_tool_ids_after_text_match_the_buffered_ids}
• gemini/tests/ir_mapping_tests.rs::reasoning_tokens_stay_inside_candidates_for_a_gemini_client
• gemini/tests/handler_tests.rs::a_duration_billed_transcription_writes_no_usage_metadata
• anthropic/tests/ir_mapping_q57_tests.rs::ant07_default_lane_drops_the_directive_as_in_1_5_5
• anthropic/tests/anthropic_hardening_tests.rs::read_response_rewrites_no_tool_use_block
• Unit golden resp_r2g_reasoning.json goes back to its bytes from before 971e2b6.

Not in this PR. L3#10's byte moves are pre-approved by the fidelity rule, but each needs a register entry plus a CHANGELOG line, and those files are outside this lane's write set (reported to the coordinator). No ledger, oracle cell or ceiling changes.

[MattJackson]

row 155 B2 (#974) edits crates/busbar-plane-llm/src/codec/{anthropic,openai_responses}/mod.rs at fn claims and the DECL quota_exceeded_status / auth_failure_status_and_kind rows, and proto_codec.rs at the DialectRef DialectCodec impl (attach_error_response_headers -> error_response_fields); disjoint from this PR's hunks.