[CompleteDotTech]

Refs #312

Slice A of #312, measurement honesty. Latency is kept in the recorder's sidecar ledger and on CallRecord only; the audit-line format and tests/request_audit.py are unchanged.

What changed, per acceptance item

Item	What	Kind
1	LiveBackend times every HTTP attempt with an injectable clock (default time.perf_counter), read immediately before and after client.system_one. Seconds are on CallRecord.latency_seconds, LiveBackend.last_latency_seconds, each last_retried entry and the sidecar lines (recorded line, retried entries, failed lines). A reading that is not a finite non-negative difference, or a clock that raises, is stored as null. Not in responses.json or the replay key.	fixed
2	vision.run_record(..., capture_date=None, n=None): keys added to a row only when given (existing rows are unchanged); a synthetic or scripted row takes n and refuses capture_date. live.load_ledger_latency(ledger_path, replay_key) returns LedgerLatency(latency_seconds, date, model) or None ("not measured").	fixed; "keys only when given" is a decision, so the existing run_record tests stay unchanged
3	vision.context_size(state, questions) returns {"characters", "utf8_bytes"} (no key naming tokens); vision.comparison_table(rows) prints not measured for every None.	fixed; the counting rule (all strings, keys and question names included, no JSON syntax, images not counted) is a decision, documented in docs/vision.md
5	docs/live.md: 64k per-request / 32k per-question and 262k with TypeSafe-Danger-Zone-Long-Context: true, source https://docs.typesafe.ai/models (read 2026-10-31); no token-counting page or endpoint in https://docs.typesafe.ai/llms.txt, so no estimator.	documented limit
6	docs/offline-and-live.md: fields holds {seed, target_characters, ...}, helpers.build_state generates the text, with a replay-key-stability test template; the same test is in tests/test_multimodal_backends.py.	documented pattern

Also documented: the latency boundaries and the ledger reader (docs/live.md), accounting rows, context size and the comparison table (docs/vision.md).

Deferred to later slices

• Item 4: a guard check that fails a synthetic or scripted notebook printing "N tokens" (needs tests/test_recipe_guard.py, which another open pull request owns).
• Item 7: the audited multi-arm pattern (with Foundation: no check that a recipe's live request count matches its README, or that no example is decided twice #213 and the Decision: may an audited notebook declare a non-live replay lane? (recipe 50, ranks 96-98, #312; #213 follow-up) #371 decision on a replay-only comparison lane).
• Item 8: the comparison helper that refuses same-model or moving-alias arms.
• The replay-only comparison lane from the Decision: may an audited notebook declare a non-live replay lane? (recipe 50, ranks 96-98, #312; #213 follow-up) #371 owner decision belongs to a later slice and is not implemented here.

Checks run (light checks only; no notebook execution)

• ruff check and ruff format --check on the five changed Python files: pass, with local ruff 0.14.0 (the repository pins 0.17.2, not installed here).
• python tools/check_hygiene.py: pass (634 files). python tools/render_catalog.py --check: pass. git diff --check: clean.
• tests/test_live.py, tests/test_vision.py, tests/test_multimodal_backends.py with test_importing_and_injecting_works_with_the_sdk_blocked deselected: 648 passed, 2 skipped. Collection listed first (650 of 651 ids; none mention notebooks or kernels).
• Each new behaviour test fails on the unmodified origin/main source (87 failures there across the three modules: the new tests, the four updated sidecar tests, and the environment failure named below) and passes here; 37 hand-made mutants of the new rules and boundaries are each killed by at least one test.
• The documentation examples were executed (the output blocks are their real output); the test template passes.

Not yet verified

• test_importing_and_injecting_works_with_the_sdk_blocked fails on unmodified origin/main too on the author's host (the package is not installed, and the test runs python -I); it was not run here.
• No SDK-gated test ran (typesafe_sdk is not installed), so what the real SDK does inside system_one is not exercised.
• The full suite, the recipe guard and every test that executes notebooks were not run.
• The limits come from the docs page as read on 2026-10-31; nothing was measured against the live API.
• Two independent Opus reviews are required (shared interface: the ledger format).

Prior work

• Foundation: image inputs and the Box primitive across backends, with local vision helpers #297 (merged): image inputs and run_record; its review noted that the ledger recorded no latency.
• Foundation: image requests are not accounted for: repeated and instruction-embedded images multiply silently and the 32 MiB aggregate caps are never checked #313 / Foundation: account for image copies and bytes, preflight the aggregate cap (#313) #372 (merged): per-call image counts in the ledger lines, which this extends with latency beside them.
• Foundation: no check that a recipe's live request count matches its README, or that no example is decided twice #213 / Foundation: declare a recipe's live request count once and audit it against the executed notebook (#213) #361 (merged): the declared request count and AuditedBackend audit lines; untouched here.
• Decision: may an audited notebook declare a non-live replay lane? (recipe 50, ranks 96-98, #312; #213 follow-up) #371 (owner decision, comment 6545320824): a replay-only comparison lane is a later slice.
• Foundation: CONTRIBUTING section 2, the honesty rule, has no mechanical check at all #218, Foundation: mechanical checks for the honesty rule, and what they cannot reach (#218) #351, Foundation: exclude the committed request-audit line from the honesty guard's quantity pool (#213, #376) #377 (honesty rule checks) and Foundation: unterminated committed output fuses a number to the next unit, so the honesty guard backs a latency or price the notebook never printed #380; Foundation: guard citations a reader cannot open and cell ids cited in prose (#225, #250) #365 (guard citations) owns the file item 4 needs.
• Searched issues and pull requests for "latency", "312", "run_record" and "context_size": no earlier implementation.

Generated with Claude Code