fix(agent): refresh DeepSeek pricing to the 2026-08-16 official rates

[Parker-Fawcett]

Partial for #94221 (point 5 — the Python-side slice; the desktop Node/ekko layer lives outside this repo).

What & why

DeepSeek changed API pricing effective 2026-08-16 and added peak-hour
billing. The pricing table still carried the retired 2026-07 snapshot, so
every DeepSeek session's cost display under-reported real spend:

model	old (table)	new official
v4-pro	0.435 / 0.87 in-out	0.66 / 1.98 off-peak
v4-flash (+ deprecated chat/reasoner aliases)	0.14 / 0.28	0.22 / 0.66 off-peak
cache-hit input	0.003625 / 0.0028	0.022 / 0.007

Verified today against DeepSeek's own pricing page
(api-docs.deepseek.com/quick_start/pricing), not third-party trackers.

Peak-hour caveat, documented in-table: DeepSeek now bills
01:00–04:00 and 06:00–10:00 UTC Mon–Fri at exactly 2× the rates above.
PricingEntry has no time-of-day axis and estimate_usage_cost sees no
timestamp, so rows carry the OFF-PEAK rates — estimates are an off-peak
floor, never a peak quote. Modeling the split would be a schema change;
flagging it as possible follow-up rather than smuggling it into a data
refresh.

Also adds the now-listed deepseek-v4-flash-vision-exp row at flash
parity (image tokens bill as input tokens per the docs), preventing the
next "unknown cost" report for that model.

How to test

from agent.usage_pricing import get_pricing_entry
e = get_pricing_entry("deepseek-v4-pro", provider="deepseek")
float(e.input_cost_per_million)   # 0.66 (was 0.435)

Updated test_deepseek_v4_pro_pricing_entry_exists to the new rates;
added test_deepseek_vision_exp_prices_as_v4_flash parity invariant; the
existing alias invariant (chat/reasoner ≡ flash) passes untouched.

Suite: scripts/run_tests.sh tests/agent/test_usage_pricing.py tests/agent/test_billing_usage.py tests/agent/test_account_usage.py → 53 passed, 0 failed at head. macOS 26, Python 3.11.

[Enough1122]

AI code review — automated review for reference, author can ignore or act on any point.

Overall: two independent fixes bundled (DeepSeek pricing refresh + hindsight background-thread secret scoping) — both look correct individually; consider splitting them in the PR description since they touch unrelated subsystems and reviewers/CI-bisect will treat them separately. Per-change notes:

1. Off-peak-floor semantics deserve user-visible surfacing — agent/usage_pricing.py:509-513: documenting "estimates are an off-peak floor" in the table comment is right, but anyone reading hermes insights cost numbers during peak hours (01:00–04:00 / 06:00–10:00 UTC Mon–Fri) will see up-to-2× understated costs with no hint. If the CostResult path can carry a caveat string cheaply, surfacing "peak hours may bill ~2×" next to deepseek rows would prevent trust erosion in the estimator; otherwise at least mirror the caveat where estimates render.

2. Misconstruction under multiplexing should warn, not debug-log — plugins/memory/hindsight/__init__.py:768-777: a provider constructed with no HERMES_HOME override while multiplexing doesn't just risk wrong credentials — self._profile_home also scopes the memory bank identity, so retains/recalls silently land in the process-default profile's bank (cross-profile memory contamination, harder to notice than a secret failure). That's a correctness hazard worth logger.warning, matching the fail-closed posture the rest of this change enforces.

3. Minor: _daemon_start_body (hindsight/__init__.py:1860) leaves open(log_path, "a") handles owned by the Rich Console unclosed on every daemon start — pre-existing, but this PR touches exactly these lines; a with open(...) as f: around the console assignment (or an explicit close after daemon start settles) is a one-line cleanup.

4. Nit: _prefetch_background re-reads self._prefetch_result fields after recall — unchanged from before, but note the thread-reference overwrite in queue_prefetch means two rapid queries race on self._prefetch_*; pre-existing, out of scope.

5. Positive: the TestBackgroundSecretScope sync-thread fixture design (making threading.Thread synchronous and capturable) tests the actual scope-install behavior rather than mocking it away — nice.

[Parker-Fawcett]

Thanks for the review — clarifying the grouping: this PR is pricing-only (DeepSeek rows + tests). The hindsight background-thread fixes you mention live on #93028 (and the debug-log vs warning question there was per Enough1122's original guidance for a debug-level diagnostic — warning would be noisy when multiplexing is off, which is the common case).

On point 1 (off-peak floor surfacing) — agreed the table comment alone won't reach users staring at hermes insights during peak. The rows here intentionally carry the off-peak floor with pricing_version deepseek-pricing-2026-08-offpeak so estimates never overstate the bill (peak is exactly 2×, per the docs). If you want a UI caveat (peak hours may bill ~2×) plumbed through CostResult, happy to add it in a follow-up — kept this PR data-only to stay focused and keep the diff reviewable. The estimator has no timestamp today, so any peak-aware math would be a schema/behavior change beyond a snapshot refresh.

[Enough1122]

Thanks for clarifying the scope split — understood that this PR is intentionally pricing-only and the hindsight background-thread fixes live on #93028. The off-peak floor rationale (estimates never overstate the bill, peak is exactly 2×, no timestamp in estimator today so peak-aware math would be a behavior change) makes sense to keep this data-only; a UI caveat via CostResult as a follow-up works. And the debug-log vs warning point being per the original guidance for a debug-level diagnostic when multiplexing is off — noted, warning would be noisy in the common case.

No further items from me on this PR.