Rework-grade translation defects: phillips_lost_conquest renders backward-looking as its antonym; sir_model inverts its closing logic

The 2026-08-20 full-corpus quality review (project-translation#50; all 143 prose pairs, mean 8.88) found three rework-grade lectures. One (hoist_failure's duplicated frontmatter) is already in flight as #265; the other two are semantic inversions that mislead the reader:

phillips_lost_conquest (scored 6.5)

backward-looking is rendered 前瞻性 (forward-looking) — the exact antonym — in all three occurrences, including inside Proposition lc_prop2, where it contradicts the same sentence's 滞后通货膨胀项 (lagged-inflation) premise. It inverts the artificial-Fed-vs-nature contrast the lecture is built on. Fix: 前瞻性 → 后顾性 (or 回顾性/后视性 per glossary preference) at all three sites:

EN: "We give the artificial Fed a backward-looking Phillips curve with drifting coefficients"
ZH: 我们赋予人工美联储一条具有漂移系数的前瞻性菲利普斯曲线

(This is the lecture that previously hit the action-translation#118 parity class and was retranslated by CLI on 2026-07-24.)

sir_model (scored 6.7)

The lecture's closing epidemiological point is reversed:

EN: "Pushing the peak of curve further into the future may reduce cumulative deaths if a vaccine is found."
ZH: 如果我们能够将感染高峰推迟到疫苗研发出来之前，就有可能降低累计死亡人数。

The Chinese says delay the peak until before a vaccine is found; the source means delay the peak so a vaccine has time to arrive first.

Rider: eight lectures carry an added figsize line

plt.rcParams["figure.figsize"] = (11, 5) appears with no source counterpart in ols, multi_hyper, samuelson, markov_perf, lq_inventories, cass_koopmans_2, sir_model, wealth_dynamics (plus an unused from numba import jit in eig_circulant). Mostly inert (plots set explicit figsize) but they are literal code deltas versus source and will confuse future diff-based syncs. Removing them touches code cells, so batch with sir_model's fix and run the cache once before the next publish tag.

Concrete quotes and locations for every finding: data/quality.json in the audit directory.

Start with data/quality.json in the audit directory, then inspect the phillips_lost_conquest and sir_model lecture translations and their source counterparts. Correct the three backward-looking occurrences and the vaccine-timing sentence, remove the listed extra figsize lines and unused numba import, and run the cache once before publishing.