RELEASE 2.0.0: version bump and changelog (merge LAST, then tag)

[AminDhouib]

The 2.0.0 version bump + CHANGELOG. Merge this last.

Human decision 2026-07-25: tag 2.0.0 now, at W5, rather than after W16 as
plan_RELEASE originally stated. The cost was stated and accepted. This PR makes that
decision honest rather than quiet — the changelog declares W6–W16 as not evidenced
instead of letting their absence read as a passing result.

Merge order (this PR is the last step before the tag)

#42 → #43 → #44 → #46 → #49 → #50      (the main line, bottom-up)
#45 → #47                               (the W4/stealth line)
#48                                     (W3 packaging, behind #46)
#52                                     (get_cookies, behind the above)
this PR                                 (version bump — LAST)

Do not merge #51. It is the staging integration branch, used only to give W5 one
tree to generate a contract from. Its content is the merge of the PRs above.

Then:

git tag v2.0.0 && git push origin v2.0.0

That runs the full three-OS gate at the tag SHA. Only if it is green does publish
download the gate-built artifact, re-check its SHA-256 against the build manifest, and
upload those exact bytes. It never rebuilds.

What changed here

File	Change
pyproject.toml	1.2.0 → 2.0.0
README.md	both install pins
RUNBOOK.md	the status example output
CHANGELOG.md	new

Deliberately left alone

• singleton.py's <= 1.2.0 comments are historical statements about which released
  backends lack the version file. They are correct as written and must not be bumped.
• tools/package_verify.py and tests/ use 1.2.0 as an illustrative sample version.
  Verified nothing binds the real package version — there is no importlib.metadata /
  __version__ assertion anywhere in src/ or tests/, and
  test_package_verify.py + test_singleton_version_aware.py pass (53/53).
• smoke_mcp.py hardcodes "version":"1.2.0" and would have broken, but it is
  gitignored (untracked local helper), so it was fixed in the working copy only.

Scope

2.0.0 carries 85 commits since the untagged 1.2.0 — the whole audit campaign, not
just the recent release work.

The breaking change is the lead item

STEALTH_MCP_SESSION_STORAGE_CAP_GB → STEALTH_MCP_BROWSER_SESSION_STORAGE_CAP_GB
and --session-cap-gb → --browser-session-cap-gb, with no back-compat alias.
Anyone setting either loses it silently on upgrade and reverts to the default. This
is the single most user-visible thing in the release, so it leads the changelog rather
than sitting in a register.

What 2.0.0 does and does not claim

Fixed, all shipping in every prior release and all invisible to the old in-process test
seam: browsers destroyed every ~2s over real stdio; list_tabs raising after any
close_tab; every navigation after a close_tab silently switching tabs and leaking
one; close_tab/switch_to_tab failing on rediscovered targets; the headless
HeadlessChrome User-Agent tell; catch-all network interception on every zero-hook
spawn.

Not claimed, stated explicitly in the changelog: macOS navigation (F-773 — neither
"works" nor "broken"; unreproduced outside hosted runners); headless
undetectability (F-774 blanks high-entropy client hints); the switch_to_tab storage
residual; any HTTP-transport claim; and every W6–W16 workstream, listed individually as
not evidenced.

A known flake exists in install-smoke (sdist Linux/X64) on Chrome cold-spawn. It may
block the tag run and need a re-run; it is recorded rather than papered over.