dream 1.2.1 — edge-case follow-up

Two edge cases from a verification pass on 1.2.0: is_header is now gated on bullets format (a leading # in Hermes sections/paragraphs is ordinary content and dedups normally again), and the --apply preflight also rejects a side-file path that is a directory. 48 tests.

Install pin: raw.githubusercontent.com/Da7-Tech/dream/v1.2.1/dream.py sha256 039e581fcf326c47720344531f66df3967e7011162d8b2200ae653122ee702ff.

dream 1.2.0 — second adversarial audit

An Opus-4.8 audit fleet surfaced 5 real defects; all fixed with regression tests (46 tests).

• --apply is now all-or-nothing: preflights the archive/journal/state paths before overwriting the live memory and aborts cleanly on a symlink/unwritable path, instead of half-completing and dying with a raw traceback (removed entries un-archived).
• Word-order-blind dedup fixed: "A calls B" vs "B calls A" were archived as duplicates; near-dups now require shared tokens in the same order, and the equal-length tie-break keeps the newer entry.
• Bullets parser no longer swallows headers: a ## header between bullet groups could be deleted as dedup side-cargo; headers are now structural (never consolidated, re-emitted in place).
• --format as the last CLI arg no longer crashes with IndexError.
• Honest measurement: the 90-day soak's budget assertion was vacuous; a dedicated budget-stress leg now forces the squeeze path and verifies the hard limit. Char accounting documented as character/codepoint-based (not "byte-for-byte").

Install pin: raw.githubusercontent.com/Da7-Tech/dream/v1.2.0/dream.py sha256 1f94b3418fc05a71791793f13ab33837974ecfb2361ffe24906afeb115ef4704.

dream 1.1.1 — audit-hardened

fsync durability in atomic writes, enforced single clock source, SHA-pinned CI with least privilege and Windows in the matrix. 40 tests + 90-day soak in CI. Install pin: raw.githubusercontent.com/Da7-Tech/dream/v1.1.1/dream.py sha256 8fd05ba9…b9a3.

dream 1.1.0 — soak-proven

First tagged release. New 90-day simulated soak (now in CI): daily churn + nightly --apply --budget 2200 against the real code through an injected clock. Verified across all 90 nights: the hard budget never breached, the newest statement of every evolving subject won, nothing was ever lost (file ∪ archive conservation), aged-file idempotence, and byte-identical full reruns. 39 unit tests.